“To increase employment, aim for productivity improvements in fields that would create more jobs. Despite tremendous productivity gains in computing and airline travel, the United States in 2020 had 11 times more programmers and 8 times more commercial airline pilots than in 1970. This growth is because programming and airline transportation were fields with what labor economists call an elastic demand. Goods with elastic demand are those where a decrease in price results in a large increase in the quantity acquired. Agriculture, on the other hand, is inelastic in the U.S., so productivity gains have reduced the number of agriculture jobs fourfold in one human lifetime (1940 to 2020). Discussions with experts in other fields will likely uncover more opportunities for AI to increase productivity. If policymakers and practitioners aim AI systems at improving productivity in elastic fields, AI can increase employment, despite public fears to the contrary. And as recent Nobelist John Jumper observed, one way to accelerate scientific progress is to improve the productivity of scientists, which is the goal of a “scientist’s aide” (see Science). Productivity gains in science from AI could prove to be extremely valuable to society [National Academies].”

The widespread adoption of large language models (LLMs) into the firm production process may significantly alter the structure of labor demand in the United States. Economists predict that about 15% of all tasks currently performed by human workers could be completed by a generative software tool including OpenAI’s ChatGPT, and 19% of the total workforce is in an occupation where at least half of the tasks involved might be performed by LLMs (Eloundou et al. 2023). Although exposure does not indicate whether an LLM might replace a human worker or raise the worker’s productivity, either outcome implies that a given output could be produced with less labor input.

When we speak about an economy that is not first‐best, we mean an economy that deviates from the Arrow‐Debreu benchmark, i.e. that exhibits market imperfections such as information problems, missing markets, price and wage rigidities which can result in aggregate demand problems, monopolies and monopsonies, and so forth. Typically, these mean that the market equilibrium is not Pareto efficient. The utilities possibilities frontier represents the maximum utility of workers, given that of entrepreneurs, taking the market failures as given. This case is illustrated in Figure 3. The initial equilibrium is Eo, but the innovation, which would have led to greater efficiency in the absence of these market imperfections, makes workers worse off—and even with costless redistributions, there is no way that both workers and entrepreneurs can be better off. 14 An example, elaborated on by Deli Gatti et al (2012a,b), were the agricultural improvements at the end of the nineteenth century and beginning of the twentieth. The result was that agricultural prices plummeted, and so too did incomes on farms and in the rural sector. But mobility is costly—moving to the urban sector required capital, and many farmers saw their capital disappear as the value of their farms decreased. Those with loans often went bankrupt. Capital market imperfections (based on information asymmetries) meant that farmers couldn’t borrow to move to the city to where the new jobs (hopefully) would be created. But as incomes in the rural sector plummeted, they couldn’t buy the goods made by the manufacturing sector. Workers in both the rural and urban sector were worse off.10 This provides at least one interpretation of the Great Depression—at least in the short run, these innovations proved Pareto inferior.

The Great Depression can be viewed as being caused by rapid pace of innovation in agriculture (see Delli Gatti et al., 2012a). Fewer workers were needed to produce the food that individuals demanded, resulting in marked decline in agriculture prices and income, leading to a decline in demand for urban products. In the late 1920’s, these effects became so large that long standing migration patterns were reversed. What might have been a Pareto improvement turned out to be an immiserizing technological change, as both those in the urban and rural sector suffered. The general result is that noted earlier: with mobility frictions and rigidities (themselves partly caused be capital market imperfections, as workers in the rural sector couldn’t obtain funds to obtain the human capital required in the urban sector and to relocate) technological change can be welfare decreasing. The economy can be caught, for an extensive period of time, in a low level equilibrium trap, with high unemployment and low output. In the case of the Great Depression, government intervention (as a by‐product of World War II) eventually enabled a successful structural transformation: The intervention was not only a Keynesian stimulus, but facilitated the move from rural farming areas to the cities where manufacturing was occurring at the time; and facilitated the retraining of the labor force, helping workers acquire the skills necessary for success in an urban manufacturing environment, which were quite different from those that ensured success in a rural, farming environment. It was, in this sense, an example of a successful industrial policy. There are clear parallels to the situation today in that a significant fraction of the workforce may not have the skills required to succeed in the age of AI.

Cognitive breadth. As suggested by the phrase “country of geniuses in a datacenter,” AI will be capable of a very wide range of human cognitive abilities—perhaps all of them. This is very different from previous technologies like mechanized farming, transportation, or even computers.39 This will make it harder for people to switch easily from jobs that are displaced to similar jobs that they would be a good fit for. For example, the general intellectual abilities required for entry-level jobs in, say, finance, consulting, and law are fairly similar, even if the specific knowledge is quite different. A technology that disrupted only one of these three would allow employees to switch to the two other close substitutes (or for undergraduates to switch majors). But disrupting all three at once (along with many other similar jobs) may be harder for people to adapt to. Furthermore, it’s not just that most existing jobs will be disrupted. That part has happened before—recall that farming was a huge percentage of employment. But farmers could switch to the relatively similar work of operating factory machines, even though that work hadn’t been common before. By contrast, AI is increasingly matching the general cognitive profile of humans, which means it will also be good at the new jobs that would ordinarily be created in response to the old ones being automated. Another way to say it is that AI isn’t a substitute for specific human jobs but rather a general labor substitute for humans.

“It establishes that, so long as AI’s microeconomic effects are driven by cost savings/productivity improvements at the task level, its macroeconomic consequences will be given by a version of Hulten’s theorem: GDP and aggregate productivity gains can be estimated by what fraction of tasks are impacted and average task-level cost savings.” ([Daron Acemoglu, p. 1]

“Using existing estimates on exposure to AI and productivity improvements at the task level, these macroeconomic effects appear nontrivial but modest—no more than a 0.71% increase in total factor productivity over 10 years.” ([Daron Acemoglu, p. 1]

Using the capital share for the entire private business sector, 0.40, this implies that GDP gains will be equal to the TFP gains multiplied by 1.66 (= 1/(1 − 0.4)). Hence taking the baseline estimate of an increase in TFP of 0.71%, I obtain a first estimate for GDP growth due to AI of 1.1% over 10 years, or taking the presence of hard tasks into account, a lower estimate of 0.92%.

“AI will have implications for the macroeconomy, productivity, wages and inequality, but all of them are very hard to predict.” ([Daron Acemoglu, p. 2]

“Some experts believe that truly transformative implications, including artificial general intelligence (AGI) enabling AI to perform essentially all human tasks, could be around the corner.” ([Daron Acemoglu, p. 2]

“Goldman Sachs (2023) predicts a 7% increase in global GDP, equivalent to $7 trillion, and a 1.5% per annum increase in US productivity growth over a 10-year period.” ([Daron Acemoglu, p. 2]

“They reckon that the total impact of AI and other automation technologies could produce up to a 1.5 − 3.4 percentage point rise in average annual GDP growth in advanced economies over the coming decade.”

“I also do not discuss how AI can have revolutionary effects by changing the process of science (a possibility illustrated by new crystal structures discovered by the Google subsidiary DeepMind and recent neural network-enabled advances in protein folding), because large-scale advances of this sort do not seem likely within the 10-year time frame and many current discussions focus on automation and task complementarities.”

“I show that when AI’s microeconomic effects are driven by cost savings (equivalently, productivity improvements) at the task level—due to either automation or task complementarities—its macroeconomic consequences will be given by a version of Hulten’s theorem: GDP and aggregate productivity gains can be estimated by what fraction of tasks are impacted and average task-level cost savings.”

New tasks created with AI can more significantly boost productivity. However, some of the new AI-generated tasks are manipulative and may have negative social value, such as deepfakes, misleading digital advertisements, addictive social media or AI-powered malicious computer attacks.

In sum, it should be clear that forecasting AI’s effects on the macroeconomy is extremely difficult and will have to be based on a number of speculative assumptions. Nevertheless, the gist of this paper is that a simple framework can discipline our thinking and forecasts, and if we take this framework and existing estimates seriously, it is difficult to arrive at very large macroeconomic gains.

“Automation corresponds to the expansion of the set of tasks that are produced by capital (including digital tools and algorithms).” ([Daron Acemoglu, p. 3]

“TFP gains over the next 10 years from AI are about 0.71%—meaning that relative to the baseline without the current suite of AI and computer vision advances, TFP will be higher by 0.71 percentage points in 10 years, or annual TFP growth will be higher by about 0.07%. This is a nontrivial, but modest effect, and certainly much less than both the revolutionary changes some are predicting and the less hyperbolic but still substantial improvements forecast by Goldman Sachs and the McKinsey Global Institute, which I discussed in the Introduction.”

Any major technology creates adjustment costs when adopted at large scale, because other organizational aspects need to evolve as well and this is typically quite costly and slow. In the context of digital technologies, Greenwood and Yorukoglu (1997) and Brynjolfsson et al. (2021), among others, have argued that productivity gains will take a J-shaped pattern, and the former paper predicts that the flat part of the J-curve lasts no less than 20 years for digital technologies. If so, the 15.4% overall cost reductions may be a significant overestimate for the next 10 years.

“Nevertheless, these considerations make me conclude that even the 0.71% increase in TFP within the next 10 years due to AI is likely to be an upper bound on this technology’s medium-run effects.”

“It is even more challenging to put numbers on the effects of new tasks. If AI helps create new tasks that increase productivity and especially contributes to the reinstatement of workers of different skill levels into the production process, its consequences can be much more positive.”

“Following its release on November 30, 2022, ChatGPT became the fastest spreading tech platform in history, reaching approximately 100 million monthly users within just two months. Its impressive features, and the greater capabilities of the newer version ChatGPT-4, released in March 2023, soon captured imaginations, both among the general public and economic commentators. Forecasts of large productivity gains have now become commonplace.”

“While there is no question that generative AI models, including ChatGPT, have impressive achievements and have great potential for beneficial economic effects, the extent of their macroeconomic consequences remains an open question.”

“They can have more modest but still notable effects on the macroeconomy by improving productivity and reducing costs in a range of tasks. Some of the forecasts have focused on these types of improvements and still produced relatively large numbers, such as a 1.5 − 3.4 percentage point per annum increase in economic growth within a 10 year horizon.”

“If AI is used to create new tasks and products, these will also add to GDP and can boost productivity growth.”

Economists have a poor track record of predicting the future. And Silicon Valley repeatedly cycles through hope and disappointment over the next big technology.

Most advanced economies now have the same problem of low productivity growth. More than any other factor, productivity—output per unit of input—determines the wealth of nations and the living standards of their people.

The first road concerns the future of economic growthwhich is largely the future of productivity growth. The US economy has been stuck with disturbingly low productivity growth for most of the past 50 years, except for a brief resurgence in the late 1990s and early 2000s (Brynjolfsson, Syverson, and Chad 2019). Most advanced economies now have the same problem of low productivity growth. More than any other factor, productivity—output per unit of input—determines the wealth of nations and the living standards of their people. With higher productivity, such problems as budget deficits, poverty reduction, health care, and the environment become far more manageable. Boosting productivity growth may be the globe’s most fundamental economic challenge.

No matter what the long‐run implications of AI are, it is clear that it has the potential to disrupt labor markets in a major way, even in the short and medium run, affecting workers across many professions and skill levels.2 The magnitude of these disruptions will depend on two important factors: the speed and the factor bias of progress in AI. On the first factor, measured productivity has increased rather slowly in recent years, even as the world seems to be captured by AI fever.3 If AI‐related innovations enter the economy at the same slow pace as suggested by recent productivity statistics, then the transition will be slower than e.g. the wave of mechanization in the 1950 – 1970s, and the resulting disruptions may not be very significant. However, there are three possible alternatives: First, some suggest that productivity is significantly under‐measured, for example because quality improvements are not accurately captured. The best available estimates suggest that this problem is limited to a few tenth of a percentage point (see e.g. the discussion in Groshen et al., 2017). Furthermore, there are also unmeasured deteriorations in productivity, e.g. declines in service quality as customer service is increasingly automated. Secondly, the aggregate implications of progress in AI may follow a delayed pattern, similar to what happened after the introduction of computers in the 1980s. Robert Solow (1987) famously quipped that “you can see the computer age everywhere but in the productivity statistics.” It was not until the 1990s that a significant rise in aggregate productivity could be detected, after sustained investment in computers and a reorganization of business practices had taken place. Third, it is of course possible that a significant discontinuity in productivity growth occurs, as suggested e.g. by proponents of a technological singularity (see e.g. Kurzweil, 2005).

过去150年中美国人民生活水平的提高，很大程度上依赖于历史上大大小小的创新。但是，对未来美国经济进步的任何考虑都必须超越创新来考察阻力，这些阻力像狂风一样正在拖累经济进步的步伐。其中，最主要的阻力是日益严重的不平等，即1970年之后最高收入阶层享有美国经济增长成果的份额不断扩大。

下一个25年商店里会有什么东西呢？技术变革会加快吗，会推动全要素生产率的增速远远高于过去40年的增速吗？全要素生产率在2004—2014年这10年中的缓慢增长是否表明，之前10年，即1994—2004年的互联网革命是自成一格且不可能复制的成就？预测未来之前，我们需要问的是这种预测的尝试是否可行。

经济史学家，包括我的同事乔尔·莫克尔通常都认为，人的大脑是无法预测未来的创新的。他毫无保留地指出：“历史总是对未来的一个坏指引，经济史学家应该避免做出预测。”他认为，工具对结果来说是必需的。例如，如果约瑟夫·李斯特没有在19世纪20年代发明消色差透镜显微镜，巴斯德就不可能发现他的细菌理论。莫克尔对未来技术进步的乐观情绪部分建立在近期出现的琳琅满目的新工具上，它们能带来进一步的研究进展，如“DNA测序仪和细胞分析”，“高性能计算机”，以及“天文、纳米化学和基因工程”。他认为，促进科技进步的核心工具之一是“令人眼花缭乱的快速搜索工具”，它们让所有的人类知识能够即时可用。

莫克尔关于未来进步的例子并不是集中关注数字化，而是包括防治传染病，减少化肥过度使用造成的环境污染所需的技术，他还发人深省地质问“新技术能阻止全球变暖吗？”应该注意的是，对抗污染和全球变暖的创新是对抗“坏的”，而不是制造“好的”。过去两个世纪的创新为消费者带来了一系列奇迹般的新商品和新服务，与这种提高生活水平的方式不同，遏制污染和全球变暖的创新是为了设法防止生活水平的下滑。

莫克尔和其他历史学家嘲笑预测未来的任何尝试；任何悲观看待未来的行为都被谴责为缺乏想象力，并且注定要重复过去悲观主义者的错误。但是他们共同的假设，即未来的创新不可预测，是错误的。有历史先例表明可以提前50年或100年做出准确的预测。在回顾分析其中一些案例之后，我们再回到今天，预测下一个25年。

Slow productivity growth is then simply a problem of not fully recognizing all the benefits we are getting from new innovations. For example, Google’s chief economist, Hal Varian, argues that slow productivity growth is rooted in mismeasurement: we are not accurately incorporating consumer benefits from products such as smartphones that simultaneously act like cameras, computers, global positioning devices, and music players. Nor are we appreciating the true productivity gains from better search engines and abundant information on the web. The chief economist of Goldman Sachs, Jan Hatzius, agrees: “We think it is more likely that the statisticians are having a harder and harder time accurately measuring productivity growth, especially in the technology sector.” He reckons that the true productivity growth of the US economy since 2000 could be several times greater than statistical agencies’ estimates. In principle, consumer and productivity benefits from new technologies should be in the TFP numbers we reported, which are based on GDP growth adjusted for changes in prices, quality, and product variety. Thus, products that significantly increase consumer welfare should translate into much higher TFP growth. In practice, of course, such adjustments are imperfect, and mismeasurement can arise. Nevertheless, these problems are unlikely to explain away the productivity slowdown. The same problem of undercounting quality improvements and broader social benefits from new products has been around ever since national income statistics were first devised. It is far from clear that digital technologies have worsened this problem. Indoor plumbing, antibiotics, and the highway system generated a panoply of new services and indirect effects that were only imperfectly measured in national statistics. Moreover, measurement problems cannot account for the current productivity slowdown; industries with greater investment in digital technologies show neither differential productivity deceleration nor any evidence of faster quality improvements than those that are less digital.

A few economists, such as Tyler Cowen and Robert Gordon, believe that this disappointing productivity performance reflects dwindling opportunities for revolutionary breakthroughs. In contrast to techno-optimists, they claim, the great innovations are behind us, and improvements from now on will be incremental, leading only to slow productivity growth.

The simple fact is that the US research and innovation portfolio has become highly imbalanced. Although more resources keep pouring into computers and electronics, almost all other manufacturing sectors are lagging. Recent research shows that new innovations appear to benefit more-productive larger firms, whereas the second- and third-tier firms are falling behind across the industrialized world, most likely because their investments in digital technologies are not paying off.

More fundamentally, productivity gains from automation may always be somewhat limited, especially compared to the introduction of new products and tasks that transform the production process, such as those in the early Ford factories. Automation is about substituting cheaper machines or algorithms for human labor, and reducing production costs by 10 or even 20 percent in a few tasks will have relatively small consequences for TFP or the efficiency of the production process. In contrast, introducing new technologies, such as electrification, novel designs, or new production tasks, has been at the root of transformative TFP gains throughout much of the twentieth century. As innovation has turned its back on boosting worker marginal productivity and creating new tasks for humans over the last forty years, it has also left many “low-hanging fruits.” One place we can get a sense of these forgone productivity opportunities is in the automobile industry. Although the introduction of robots and specialized software has increased output per worker in the industry, there is evidence that investing more in humans would have boosted productivity by more. This is what Japanese car companies, such as Toyota, discovered starting in the 1980s. When they automated more and more tasks, they saw that productivity was not increasing by much because, without the workers in the loop, they were losing flexibility and the ability to adapt to changes in demand and production conditions. In response, the company took a step back and reinstated workers’ central role in crucial production tasks.

Reassuringly, AI does not appear to be advancing so much that it will create mass joblessness. Like the industrial robots we discussed in Chapter 8, current technology thus far can perform only a small set of tasks, and its impact on employment is limited. Nevertheless, it is heading in a direction that is biased against workers and is destroying some jobs. Its most major likely impact is to further lower wages for many people, not create a completely workless future. The problem is that although AI fails in most of what it promises, it still manages to reduce the demand for workers.