关于实际工资，图8.7提供了一个更全面的历史角度，它描述了非管理类生产工人每小时实际工资与实际GDP之间的关系，这两者都用指数来表达，1940年等于100。这幅图最引人注目的方面是，1940年之前实际工资增长速度超过时均产出（即劳动生产率），但后来增长速度很慢，尤其在1980年之后。1980年后的这种分化现象与大量文献研究是一致的：1980年之后不平等扩大，劳动者收入份额下降，很大一部分生产率收益流向收入分配的顶层1%。

更令人困惑的是，1940年之前实际工资相对于生产率快速增长，特别是在1920—1940年实际工资激增，尽管大萧条年代劳动力需求下降。这期间的工资增长率如图8.8所示。结果相当引人注目。在1870年至1940年的70年间，与劳动生产率相比，实际工资增长率几乎高出1个百分点，年增长率分别为1.51%和2.48%。但1940年之后，这个关系倒了过来，生产率增长2.25%，实际工资增长1.56%，两者之间的差异达到0.69个百分点。

实际工资增长最快的阶段是1910—1940年，增长率为3.08%。比1870—1910年高1个百分点。关于这一变化，一个合理的解释是早期无限制移民压低了实际工资，而“一战”终结了大规模移民，美国在1921年通过了反移民配额限制，且在1924年和1929年进一步收紧配额。罗斯福新政立法对建立工会的鼓励也同样重要。结果，在1936—1940年实际工资增长率迅速提高，达到4.64%。

As these technologies spread, many relatively high-wage occupations started declining. In 1970 about 33 percent of American women were in clerical jobs, which paid decent salaries. Over the next six decades, this number declined steadily and is now down to 19 percent. Recent research documents that these automation trends have been a powerful contributor to the wage stagnation and declines for low- and middle-skill office workers.

“First, our model imposes that it is always the tasks at the bottom that are automated; in reality, it may be those in the middle (e.g., Acemoglu and Autor 2001). Incorporating the possibility of such “middling tasks” being automated is an important generalization, though ensuring a pattern of productivity growth consistent with balanced growth in this case is more challenging.”

Across ten different advanced and emerging economies, a majority of citizens believe that automation will replace existing jobs, increase inequality, and make it harder to find work (Wike and Stokes 2018). Yet automation anxiety is nothing new. In 1964, President Lyndon Johnson created the Blue Ribbon National Commission on Technology, Automation, and Economic Progress, which concluded (among other things) that the US should have a guaranteed minimum income to support workers through the coming wave of automation. Of course, that wave did not come to pass, and the employment-to-population ratio is substantially higher now than it was in 1964. Is this time different?

First, we show that job polarization—meaning employment growth at the bottom and top of the wage distribution and declines in the middle—is no longer an accurate description of what is happening in the US labor market. In recent years, employment has declined in both low- and middle-paid occupations and grown rapidly in high-paying managerial, professional, and technical jobs. Skill upgrading describes what is happening in the post-pandemic US labor market better than polarization does.

Second, we show that the growth of low-paid service jobs like home health aides, food preparation and service workers, cleaners, barbers, and fitness instructors has stalled completely since 2010, after having grown rapidly in the 1990s and the first decade of the 2000s (Autor and Dorn 2013). Service employment cratered during the COVID-19 pandemic, but employment growth in service occupations slowed in the early 2010s, likely because of rising labor costs (Autor, Dube, and McGrew 2023).

Third, we find that science, technology, engineering, and math (STEM) employment has grown rapidly over the last decade. STEM employment grew from 6.5 percent of all jobs in 2010 to nearly 10 percent in 2024, an increase of more than 50 percent. The growth in STEM work is concentrated in occupations like software developers and programmers, although we also see increases across a broad range of science and engineering occupations. STEM job growth has accelerated especially rapidly since 2017 and is matched by increasing private-capital investment in AI-related technology.

Our first key finding is that we uncover substantial declines in employment for early-career workers (ages 22-25) in occupations most exposed to AI, such as software developers and customer service representatives. In contrast, employment trends for more experienced workers in the same occupations, and workers of all ages in less-exposed occupations such as nursing aides, have remained stable or continued to grow.

Our second key fact is that overall employment continues to grow robustly, but employment growth for young workers in particular has been stagnant since late 2022. In jobs less exposed to AI young workers have experienced comparable employment growth to older workers. In contrast, workers aged 22 to 25 have experienced a 6% decline in employment from late 2022 to July 2025 in the most AI-exposed occupations, compared to a 6-9% increase for older workers. These results suggest that declining employment AI-exposed jobs is driving tepid overall employment growth for 22- to 25- year-olds as employment for older workers continues to grow.

Our third key fact is that not all uses of AI are associated with declines in employment. In particular, entry-level employment has declined in applications of AI that automate work, but not those that most augment it. We distinguish between automation and augmentation empirically using estimates of the extent to which observed queries to Claude, the LLM, substitute or complement for the tasks in that occupation. While we find employment declines for young workers in occupations where AI primarily automates work, we find employment growth in occupations in which AI use is most augmentative. These findings are consistent with automative uses of AI substituting for labor while augmentative uses do not.

Fifth, the labor market adjustments are visible in employment more than compensation. In contrast to our findings for employment, we find little difference in annual salary trends by age or exposure quintile, suggesting possible wage stickiness. If so, AI may have larger effects on employment than on wages, at least initially.

Fourth, we find suggestive evidence of AI-related employment disruption in retail sales. While large language models (LLMs) like ChatGPT are too new for us yet to see any direct impact on the labor market, companies have been using predictive AI to optimize business operations since at least the mid-2010s. Online retailers like Amazon use AI to personalize prices and product recommendations and to manage inventory more efficiently, outcompeting big-box retail (see, for example, Deming 2020). We find that retail sales jobs have declined by 25 percent in the last decade. There were 850,000 fewer retail sales workers in the US in 2023 compared to 2013, even though the US economy added more than 19 million jobs over this period. The decline in retail sales began long before COVID-19 but has accelerated in the last few years. Labor productivity growth in retail sales has also outpaced that in other sectors. Interestingly, online retail has also led to growth of “last-mile” jobs like lighttruck delivery drivers and stockers and order fillers. The decline of retail sales fits a broader pattern of technology-fueled occupational upgrading in white-collar office work. Since 1990, front-office jobs like secretaries and administrative assistants and back-office jobs like billing and financial processing have declined by more than 50 percent as a share of all jobs in the US economy. Yet managers and business analysis jobs have grown rapidly over this same period. While this change may partly reflect title inflation (for example, manager versus supervisor), it also captures a shift away from routine monitoring and categorizing and toward strategy and decision-making (Deming 2021). Our best guess is that this trend will continue. We expect continued declines in routine office work and retail sales jobs, and a ratcheting up of firms’ expectations of managers and business analysts, who will now be valued only to the extent that they can use AI to become more productive. Taken together, these four facts suggest that we may be entering a period of more pronounced labor market disruption. To illustrate this point, we recalculate our measure of labor market “churn” through 2022. Although the 2010s were very stable pre-pandemic, the post-pandemic labor market has changed dramatically. A key outstanding question is whether the labor market disruption of the past few years is a temporary response to the changes wrought by COVID-19 or an early sign of AIfueled labor market disruption.

Automation technologies, including specialized software tools, computerized production equipment, and industrial robots, have been spreading rapidly throughout the industrialized world. For example, the share of information processing equipment and software in overall investment in the US has increased from 3.5% to over 23% between 1950 and 2020 (BEA, 2021a), while the number of industrial robots per thousand workers has risen from 0.38 in 1993 to about 1.8 in 2017 (BEA, 2021b; IFR, 2018). There is growing evidence that these technologies have not just automated a range of tasks previously performed by workers and impacted the wage structure,1 but also have led to polarization of employment and wages—meaning that the negative effects have concentrated on employment and wages in the middle of the wage distribution.2 This pattern is intimately linked to the fact that many of the tasks that have been automated used to be performed by middle-skill workers.

“Polanyi’s paradox”， Michael Polanyi’s statement that “we can know more than we can tell”， Put simply, many of the manual and abstract tasks embed rich tacit knowledge, making them non-routine. Because routine tasks are technologically easier to automate and are performed by middle-skill workers located in the middle of the wage distribution, new automation technologies have displaced labor from middle-skill occupations and have had their most negative effects on middle-pay worker groups.

automation has focused on middle-skill tasks, because these are the most profitable ones to automate. Specifically, low-skill tasks can be performed at lower labor expenses, reducing the cost advantage of machines relative to humans.

Employment polarization here simply means that human workers are squeezed into smaller sets of tasks at the bottom and the top. Wage polarization takes a more specific form: relative wage changes increase as a function of the distance between the task that a skill type performs and the boundaries of the set of automated tasks. As a result, we prove that skill premia increase among worker types performing more complex tasks than those that are automated and decrease among worker types performing less complex tasks than the automated ones. Put differently, interior automation hurts (relatively) workers that are closer to the set of automated tasks. This is intuitive in view of the fact that workers closer to this set used to have a stronger comparative advantage for tasks that are now automated.

Fourth, we use the model to study global—as opposed to local—effects of automation, which result when there are large declines in costs of capital goods. We show that as long as these changes keep us in the region of interior automation, their effects are qualitatively the same as those of local changes. Ultimately, however, automation expands from the interior of the set of tasks to take over all lowskill tasks. When this happens, the pattern of polarization reverses. While an expansion in interior automation hurts workers in the middle of the skill distribution the most (and lowest-skill workers are to some degree sheltered), a switch from interior to low-skill automation has its most adverse effects on lowest-skill workers. Hence, our model predicts that as automation proceeds, its inequality implications may become worse, not just quantitatively but also qualitatively.7

底层90%群体的工资下行压力

哪些因素影响了第90百分位之下的收入分配的演变？20世纪70年代中期标志着一个转折点的到来，从一个收入分配中层和底层的工资稳步上升的时代转向一个新时代，也就是过去40年底层人群工资增长很少而顶层人群收入迅速增长。在过去30年里，美国工资的巨大停滞导致许多观察者认为美国经济基本上走向衰落。那么，是什么导致了40年前的逆转？

1929—1945年，顶层收入增长比底层和中层慢一些，形成了向更平等转变的趋势，这被克劳迪娅·戈尔丁和罗伯特·马戈称为“大压缩”。“二战”至1975年，收入顶层和底层的增长率大致相同，因此大压缩历时约30年。戈尔丁和马戈强调，支撑大压缩的因素有三个，即工会力量上升、贸易下降和移民减少。这三个因素可以追溯到20世纪30年代，令人信服地解释了在1945—1975年不平等为什么处于较低水平，而且三个因素的反向变化成为1975年之后不平等加剧的重要原因之一。在这一小节中，我们研究这种反向变化，即工会力量下降、进口上升和移民增加，同时研究其他两个通常被认为加剧了不平等的因素，即自动化和最低实际工资下降。随后，我们转过来分析教育对不平等加剧的作用，因为大学毕业生的工资增长与高中毕业生和辍学生的工资停滞形成鲜明对比。

加入工会的美国员工比例迅速下降，从1973年的27%下降到1986年的19%，然后再慢慢滑落至2011年的13%。工会参与率的下降导致工资减少，特别是工资中位数。在工会参与率下降的同时，市场力量在增强，尤其是制造业工作岗位缩减，消费需求由商品转向服务，还有许多雇主采取强势的反工会立场。由于越来越多的企业，特别是制造业企业和建筑业企业从临时的职业介绍机构雇用工人，发放相对较低的工资，提供很低的额外福利，所以工资停滞甚至下降之势加速。企业不仅受益于较低的劳动力成本，而且受益于弹性增加，可以根据需求调整工人的工作时间。

美国GDP中的进口份额从1970年的5.4%增加到2014年的16.5%。进口货物中包含的劳动力是对国内劳动力的替代。因此，GDP中进口份额增加导致非技能工人和中等技能工人相对工资出现下降。在一份特别引人注目的分析中，戴维·奥托及其合著者的计算指出，1990—2007年从中国的进口量约占制造业就业下降量的1/4，这既降低了工资，又降低了劳动参与率，还增加了公共财政的转移支付。进口货物的侵袭并不限于最终产品，因为企业和国家越来越专注于不同的生产阶段。例如，2001—2014年汽车零部件进口从630亿美元增加到1380亿美元，增加一倍多，致使美国许多零部件制造商关闭国内工厂，而且在某些情况下，零部件“离岸”到国外生产，尤其是墨西哥。综上所述，进口渗透率和外包增长代表了全球化对国内就业和工资水平的综合影响。在上述汽车零部件行业的案例中，全球化的影响还包括工资中位数从2003年的每小时18.35美元下降至2013年的15.83美元。

1995—2005年的10年间，移民占美国总劳动力增长的一半以上。作为一个补充指标，可以看到劳动力中外国出生的工人比例由1970年的5.3%稳步上升至2005年的14.7%。经济研究显示，移民造成国内工人的工资小幅下降，但对缺乏高中教育的国内工人影响最大。许多低技能移民找到并不匹配的工作，或进入已经由国外出生的工人构成的职业，例如餐厅服务和导游服务，因此他们的主要影响是降低了国外出生的工人的工资而不是国内出生的工人的工资。以往的文献指出，在高中辍学者中，1980年以前国内和国外出生的工人几乎有相同的工资，但2004年以来国外出生的工人的工资降低了15%~20%。

即使工会没有变弱，进口和移民没有增加，底层90%群体的工资下行压力也可能出现。自动化的稳步推进、机器替代人工，将促使底层90%群体的相对收入下降。薪水相对较高的制造业工作岗位已经被侵蚀，因为美国制造业就业份额从1953年的30%降至目前的不到10%。自动化影响与“技能偏向的技术进步”叠加，致使软件驱动的计算机替代了日常工作，而这些工作的损失不只发生在制造工厂的装配线上，还发生在打字员、簿记员、文员、接待员等日常办公职业中。自动化并没有造成悲观主义者曾经担心的大规模持久失业，在2007年底结束的经济周期扩张期，美国经济的失业率也能控制在5%以内，并且在2015年失业率再次下降至近5%。

尽管就业岗位没有大规模消失，但就业构成发生了变化，在职业分布的顶层和底层创造出更多工作机会，中层则出现空心化。这种转变被称为“极化假说”，在最近几年已被劳动经济学家广泛证明。管理人员和专业人士所从事的上层工作，通常被称为“非常规”职业。由流水线生产工人、簿记员、接待员和职员所从事的中层工作被称为“常规”职业，而底层工作被称为“体力活”。中等技能型常规工作岗位消失造成的一个后果是，中等技能工人被迫参与竞争低技能体力劳动职业，从而形成了体力劳动工人供大于求。其结果是高中毕业生和高中辍学生等低技能工人的工资下降，如图18.3所示。一旦就业构成中相对高薪的制造业工作岗位转向大范围的典型低工资岗位，诸如零售、食品、清洁服务、草地清理工作，总体工资水平也就随之降低。

Why might AI adversely affect exposed entry-level workers more than other age groups? One possibility is that, by nature of the model training process, AI replaces codified knowledge, the “book-learning” that forms the core of formal education. AI may be less capable of replacing tacit knowledge, the idiosyncratic tips and tricks that accumulate with experience. 5 As young workers supply relatively more codified knowledge than tacit knowledge, they may face greater task replacement from AI in exposed occupations, leading to greater employment reallocation (Acemoglu and Autor, 2011). In contrast older workers with accumulated tacit knowledge may face less task replacement. These benefits of tacit knowledge may accrue less to non-college workers in occupations with low returns to experience. Furthermore, more experienced workers may be more skilled in other ways, making them less vulnerable to substitution by AI tools (Ide, 2025). An important direction for research is to further model and test these predictions.

“This calculation implies that 19.9% of US labor tasks are exposed to AI.”

“I take the average labor cost savings to be 27%—the average of the estimates in Noy and Zhang (2023) and Brynjolfsson et al. (2023)—and turn this into total cost savings using industry labor shares, which imply an average total cost savings of 15.4%.”

For each age group, employment growth from late 2022 to July 2025 was 6-13% for the lowest three AI exposure quintiles, with no clear ordering in employment growth by age. In contrast, for the highest two exposure quintiles employment for 22-25 year olds declined by 6% between late 2022 and July 2025, while employment for workers aged 35-49 grew by over 9%. These results show that declining employment in AI-exposed jobs is driving tepid overall emplyoment growth for workers between the ages of 22 and 25.

In fact, from equation (11), abstracting from the productivity effect, the direct impact on the equilibrium wage will be a (σ − 1)/σ% change. When σ < 1, which is the plausible case as discussed above, this is negative. The overall impact may still be an increase in wages because of the productivity effect, but as already noted, when σ is approximately equal to the share of capital in national income, the overall impact will be essentially zero. In conclusion, without the creation of a sufficient number of new tasks, inequality between capital and labor will increase and wage rises may be limited.

Although the abatement of rent sharing and the automation focus of new technologies have been the most important drivers of inequality and the decline of the labor share, other factors have also played a role. Offshoring has contributed to worsening conditions for labor: numerous jobs in car manufacturing and electronics have been shifted to lower-wage economies, such as China or Mexico. Even more important has been rising merchandise imports from China that have adversely affected many US manufacturing industries and the communities in which they were concentrated. The total number of jobs lost to Chinese competition between 1990 and 2007, just before the Great Recession, may be as high as three million. However, the effects of automation technologies and the eclipse of rent sharing on inequality have been even more extensive than the consequences of this “China shock.” Import competition from China impacted mostly low-value-added manufacturing sectors, such as textiles, apparel, and toys. Automation, on the other hand, has concentrated in higher-value-added and higher-wage manufacturing sectors, such as cars, electronics, metals, chemicals, and office work. It is the dwindling of this latter set of jobs that has played a more central role in the surge in inequality. As a result, although competition from China and other low-wage countries has reduced overall manufacturing employment and depressed wage growth, it has been the direction of technological change that has been the major driver of wage inequality.

The evolution of federal research and science policy may have been another contributing factor. Starting from before World War II, government funding of science and private-sector research was generous, especially in areas that were national defense priorities. This was a powerful inducement to new critical areas, such as antibiotics, semiconductors, satellites, aerospace, sensors, and the internet.

Business-As-Usual Scenario: This scenario assumes that task complexity follows a Pareto distribution with an infinite right tail, and the fraction of non-automated tasks declines at a constant rate. This implies that there are always tasks that only humans can perform or that we choose not to automate. Under our assumptions, output grows steadily at about 2 percent per year. For our parameterization, both the returns to capital and the wage bill rise indefinitely, approximately in tandem. More generally, there is a race between automation and capital accumulation – automation tends to pull down wages whereas the accumulation of capital that is complementary to the remaining tasks for labor increases wages. If automation proceeds too quickly, wages may decline, even if full automation is never reached. 2. Baseline AGI Scenario: In this scenario, the task complexity of human work follows a distribution with a finite upper bound, and full automation is reached within 20 years, in line with Hinton’s upper range for when AGI will be reached. The economic implications are stark: wages grow during the initial period but collapse before full automation is achieved. After the wage collapse, labor and compute earn equal returns. The economy transitions to a steadystate growth of 18 percent per year, but the labor share declines dramatically as automation progresses. 3. Aggressive AGI Scenario: This scenario is similar to the Baseline AGI scenario but assumes a shorter-tailed distribution, with full automation reached within 5 years, in line with Hinton’s shorter estimate for AGI. The wage collapse starts much earlier, in year 3. The economy experiences a more rapid transition to high growth rates, and labor becomes perfectly substitutable with compute very quickly. This scenario results in a quick concentration of returns to capital owners. 4. Bout of Automation Scenario: The final scenario assumes that the economy experiences a large bout of automation in the short run, for example, because cognitive labor is automated, but that there remains an infinite tail of tasks that either cannot be automated or that we choose not to automate. Initially, rapid automation leads to a wage collapse similar to the Aggressive AGI scenario. However, as the economy accumulates more capital, labor becomes scarcer again. Eventually, wages rise above the return to capital and start growing again. This scenario illustrates the possibility of a temporary labor demand collapse followed by recovery. These scenarios highlight the critical role of (i) the complexity required to automate different work tasks and (ii) the speed of progress in AI in determining the dynamics of output and wages during the transition to AGI and beyond. They underscore the importance of understanding the nature of human cognitive capabilities and any potential limits of machine intelligence in shaping our economic future. The stark differences between the scenarios also emphasize the importance for economic policy to be adaptive as we navigate the uncertain path towards AGI. Some argue that we need not be concerned about the fate of labor since the comparative advantages of humans and machines will always allow for gainful trade between the two (see, e.g., Smith, 2024). Although it is true that agents with different capabilities – like humans and robots will have comparative advantages, this does not prevent wage declines from technological changes. Comparative advantage just tells us that two agents with different endowments or capabilities benefit from trade – it does not tell us how these benefits will evolve over time. Under AGI, the relative terms-of-trade of humans may deteriorate significantly and may imply significant income losses for the average worker. As an analogy, horses and humans have always had comparative advantages, and they still do, yet the combustion engine has greatly undermined the economic value of horses.

The caveat is that speed of adjustment matters. The loss of jobs from the manufacturing shock largely occurred in two bursts: the back-to-back recessions of 1980-82 (-1.6 million manufacturing jobs; -8%) and the 2000-2012 China shock and financial crisis (-5.3 million manufacturing jobs, -30%). The rapid loss of manufacturing jobs was not the only factor in today’s political polarization but it was a central factor both in the U.S. and globally (Autor et al. 2020a; Hill 2021; Guriev & Papaioannou 2022). Economic adaptation depends in part on societal adaptation. For this reason, the political economist Karl Polanyi argued that a central responsibility of government is to regulate the pace of change, asserting that the speed of adoption determined the social impact: “whether the dispossessed could adjust themselves to changed conditions without fatally damaging their substance, human and economic, physical and moral; whether they would find new employment in the fields of opportunity indirectly connected with the change; and whether the effects of increased imports induced by increased exports would enable those who lost their employment through the change to find new sources of sustenance” (Polanyi 1944). Word-processing software was a shock to the typist occupation but the pace of adoption was slowed by users having to purchase computers and install the software. The number of persons employed as “Word Processors and Typists” declined from one million in 1980 to 26 33,000 today but there was no huge disruption because the decline was spread over 44 years.22 David (1990) points to similar constraints on electrification in the early 1900s. Compared to these earlier episodes, the dissemination of LLM-based services through the web and the cloud, absent regulation, will be much faster and, like electrification, very broad.

“I predicted that AI could displace half of all entry-level white collar jobs in the next 1–5 years, even as it accelerates economic growth and scientific progress.”

labor market displacement, and concentration of economic power.

There are two specific problems I am worried about: labor market displacement, and concentration of economic power. Let’s start with the first one. This is a topic that I warned about very publicly in 2025, where I predicted that AI could displace half of all entry-level white collar jobs in the next 1–5 years, even as it accelerates economic growth and scientific progress. This warning started a public debate about the topic. Many CEOs, technologists, and economists agreed with me, but others assumed I was falling prey to a “lump of labor” fallacy and didn’t know how labor markets worked, and some didn’t see the 1–5-year time range and thought I was claiming AI is displacing jobs right now (which I agree it is likely not). So it is worth going through in detail why I am worried about labor displacement, to clear up these misunderstandings.

Prominent Silicon Valley AI figures, such as Marc Andreessen, use the Lump of Labor fallacy to argue that workers shouldn’t be afraid of the new wave of AI automation either.

The false assumption is that the newly created jobs and the newly created demand for work can only be satisfied by humans. That is false. Machines can just as well fill those positions.

The Lump of Labor Fallacy will remain true: New tech will create new jobs, but humans will no longer need to fill them.

AI replacing jobs transfers wealth from knowledge workers to AI company owners. Our world is already very unequal — in the US, the richest 1% have as much wealth as the bottom 90% combined. Think about what would happen if the 79% of Americans who rely on wages lost their income and that money too went to the already wealthy instead.

Slicing by cognitive ability. Across a wide range of tasks, AI appears to be advancing from the bottom of the ability ladder to the top. For example, in coding our models have proceeded from the level of “a mediocre coder” to “a strong coder” to “a very strong coder.”40 We are now starting to see the same progression in white-collar work in general. We are thus at risk of a situation where, instead of affecting people with specific skills or in specific professions (who can adapt by retraining), AI is affecting people with certain intrinsic cognitive properties, namely lower intellectual ability (which is harder to change). It is not clear where these people will go or what they will do, and I am concerned that they could form an unemployed or very-low-wage “underclass.” To be clear, things somewhat like this have happened before—for example, computers and the internet are believed by some economists to represent “skill-biased technological change.” But this skill biasing was both not as extreme as what I expect to see with AI, and is believed to have contributed to an increase in wage inequality,41 so it is not exactly a reassuring precedent.

Ability to fill in the gaps. The way human jobs often adjust in the face of new technology is that there are many aspects to the job, and the new technology, even if it appears to directly replace humans, often has gaps in it. If someone invents a machine to make widgets, humans may still have to load raw material into the machine. Even if that takes only 1% as much effort as making the widgets manually, human workers can simply make 100x more widgets. But AI, in addition to being a rapidly advancing technology, is also a rapidly adapting technology. During every model release, AI companies carefully measure what the model is good at and what it isn’t, and customers also provide such information after the launch. Weaknesses can be addressed by collecting tasks that embody the current gap, and training on them for the next model. Early in generative AI, users noticed that AI systems had certain weaknesses (such as AI image models generating hands with the wrong number of fingers) and many assumed these weaknesses were inherent to the technology. If they were, it would limit job disruption. But pretty much every such weakness gets addressed quickly— often, within just a few months.

It’s worth addressing common points of skepticism. First, there is the argument that economic diffusion will be slow, such that even if the underlying technology is capable of doing most human labor, the actual application of it across the economy may be much slower (for example in industries that are far from the AI industry and slow to adopt). Slow diffusion of technology is definitely real—I talk to people from a wide variety of enterprises, and there are places where the adoption of AI will take years. That’s why my prediction for 50% of entry level white collar jobs being disrupted is 1–5 years, even though I suspect we’ll have powerful AI (which would be, technologically speaking, enough to do most or all jobs, not just entry level) in much less than 5 years. But diffusion effects merely buy us time. And I am not confident they will be as slow as people predict. Enterprise AI adoption is growing at rates much faster than any previous technology, largely on the pure strength of the technology itself. Also, even if traditional enterprises are slow to adopt new technology, startups will spring up to serve as “glue” and make the adoption easier. If that doesn’t work, the startups may simply disrupt the incumbents directly.

That could lead to a world where it isn’t so much that specific jobs are disrupted as it is that large enterprises are disrupted in general and replaced with much less labor-intensive startups. This could also lead to a world of “geographic inequality,” where an increasing fraction of the world’s wealth is concentrated in Silicon Valley, which becomes its own economy running at a different speed than the rest of the world and leaving it behind. All of these outcomes would be great for economic growth—but not so great for the labor market or those who are left behind.

Second, some people say that human jobs will move to the physical world, which avoids the whole category of “cognitive labor” where AI is progressing so rapidly. I am not sure how safe this is, either. A lot of physical labor is already being done by machines (e.g., manufacturing) or will soon be done by machines (e.g., driving). Also, sufficiently powerful AI will be able to accelerate the development of robots, and then control those robots in the physical world. It may buy some time (which is a good thing), but I’m worried it won’t buy much. And even if the disruption was limited only to cognitive tasks, it would still be an unprecedentedly large and rapid disruption.

Fourth, some may argue that comparative advantage will still protect humans. Under the law of comparative advantage, even if AI is better than humans at everything, any relative differences between the human and AI profile of skills creates a basis of trade and specialization between humans and AI. The problem is that if AIs are literally thousands of times more productive than humans, this logic starts to break down. Even tiny transaction costs could make it not worth it for AI to trade with humans. And human wages may be very low, even if they technically have something to offer.

Third, perhaps some tasks inherently require or greatly benefit from a human touch. I’m a little more uncertain about this one, but I’m still skeptical that it will be enough to offset the bulk of the impacts I described above. AI is already widely used for customer service. Many people report that it is easier to talk to AI about their personal problems than to talk to a therapist—that the AI is more patient. When my sister was struggling with medical problems during a pregnancy, she felt she wasn’t getting the answers or support she needed from her care providers, and she found Claude to have a better bedside manner (as well as succeeding better at diagnosing the problem). I’m sure there are some tasks for which a human touch really is important, but I’m not sure how many—and here we’re talking about finding work for nearly everyone in the labor market.

“Are such large effects plausible? And if there are going to be productivity gains, who will be their beneficiary?” ([Daron Acemoglu, p. 2]

“With previous automation technologies, such as robotics, most gains” ([Daron Acemoglu, p. 2]

“accrued to firm owners and managers, while workers in impacted occupations experienced negative outcomes” ([Daron Acemoglu, p. 3]

I also explore AI’s wage and inequality effects. My framework implies that productivity gains from AI are unlikely to lead to sizable wage rises. Moreover, even if AI improves the productivity of low- and middle-performing workers (or workers with limited expertise in complex tasks), I argue that this may not translate into lower inequality. In fact, I show by means of a simple example how an increase in the productivity of low-skill workers in certain tasks can lead to higher rather than lower inequality. Adapting the general equilibrium estimates from Acemoglu and Restrepo (2022) to the setting of AI, I find that the more intensive use of AI is unlikely to lead to substantial wage declines for affected groups, because AI-exposed tasks are more evenly distributed across demographic groups than were the tasks exposed to earlier waves of automation. Nevertheless, I estimate that AI will not reduce inequality and is likely to have a negative effect on the real earnings of low-education women (especially white, native-born women). My findings also suggest that AI will further expand the gap between capital and labor income as a whole.

Finally, I argue that as originally suggested in Acemoglu and Restrepo (2018), more favorable wage and inequality effects, as well as more sizable productivity benefits, will likely depend on the creation of new tasks for workers in general and especially for middle and low-pay workers.

In 1930, Keynes wrote an essay on the “Economic Possibilities of our Grandchildren,” in which he described how technological possibilities may translate into utility possibilities. He worried about the quality of life that would emerge in a world with excess leisure. And he thought all individuals might face that quandary. But what has happened in recent years has raised another possibility: innovation could lead to a few very rich individuals—who may face this challenge—whereas the vast majority of ordinary workers may be left behind, with wages far below what they were at the peak of the industrial age.

AGI will disrupt labor markets and could lead to unprecedented levels of income concentration. Compensating the losers would necessitate new ways of distributing income and economic benefits in society. As AGI systems become capable of performing a wide range of cognitive tasks, they could lead to widespread labor displacement, mass unemployment, and stark wage declines (Korinek and Juelfs, 2024). If left to the market, the benefits of AGI would accrue primarily to those who own capital and control AGI technologies. Our current systems of social insurance revolve primarily around work for example, people receive retirement benefits after having worked for several decades, or they receive unemployment if they lose their jobs. In many countries, health benefits are linked to work. In an AGI future in which labor markets are disrupted, compensating the losers would require new mechanisms for income distribution that are independent of work, such as, for example, Universal Basic Income (UBI). Ultimately, the challenge lies in harnessing the immense potential of AGI to create a more prosperous society for all, rather than allowing it to widen economic divides. However, implementing such systems would require significant changes to current economic models. The challenge of income inequality in the face of AI-driven technological change becomes even more daunting when we consider between-country rather than within-country disparities. As Korinek and Stiglitz (2021) argue, the impact of AI and related technologies on developing countries could be particularly severe. These countries often rely on their comparative advantage in laborintensive industries, which may be devalued by AI-driven automation. Unlike within-country redistribution, where national governments have mechanisms to compensate losers, there are no well-established global institutions for large-scale redistribution across countries. The potential for AI to exacerbate global inequality is therefore much greater, and the policy challenges far more complex. Addressing this issue would require unprecedented levels of international cooperation and potentially the development of new global economic governance structures to share the benefits of AI more equitably across nations.

AGI-induced labor market disruption could lead to widespread economic discontent, social unrest, and political instability, potentially undermining democratic institutions. If AGI leads to widespread labor displacement and stark wage declines, it may create economic insecurity for large segments of the population while also depriving workers of an important element of purpose. This disruption could have profound implications for political stability (Bell and Korinek, 2023). Historically, significant economic displacements have often led to social unrest and the rise of populist or authoritarian movements. In an AGI world, the concentration of economic power in the hands of those who control AGI technologies could exacerbate inequalities, potentially undermining democratic processes and institutions. Moreover, societies may struggle to find new ways to effectively ensure economic security and social inclusion for all. This could lead to increased polarization and conflict between those benefiting from AGI and those left behind. Additionally, the rapid pace of change brought about by AGI could outstrip the ability of social and political institutions to adapt, eroding public trust in democratic systems. Addressing these challenges would require policies to mitigate economic disruptions, ensure equitable distribution of AGI benefits, and strengthen democratic institutions to withstand the pressures of rapid technological change.

As discussed in the Introduction, a number of commentators and experts are cautiously optimistic that advances in generative AI could be beneficial for labor or at the very least not impact workers as adversely as previous waves of digital technologies, such as robotics and software systems, which were predominantly used for automation. There are three potential pathways via which such optimism may be realized. 1. AI can enable productivity increases in tasks currently produced by labor. This is the task-complementarities channel and can be captured either by an increase in AL or an increase in γL(z) for a subset of the tasks that are automated or by increases in λU and λH (which, recall, are the productivities of unskilled and highly-skilled workers). However, recall that when σ < 1, these types of productivity improvements will reduce the labor share, and thus inequality between capital and labor will increase. 2. If AI generates very large productivity gains, it may increase wages even though it reduces the labor share (Acemoglu and Restrepo, 2018, 2019b). This channel thus critically hinges on the magnitude of the productivity effects discussed above, but in any case, always increases inequality between capital and labor. 3. As already discussed, some early studies show that within narrow occupations, lowerperforming or lower-expertise workers are the ones benefiting from generative AI. This raises the possibility that AI could be more complementary to lower-skill workers and may reduce labor income inequality. In my framework, this would be captured by an increase in λU relative to λH. However, even in this case, inequality between capital and labor will rise (provided that σ < 1).

4. If AI created new (good) tasks, these would reduce inequality between capital and labor, and if enough new tasks were targeting lower-skill workers, this could also reduce labor income inequality (Acemoglu and Restrepo, 2018).

If AI created new (good) tasks, these would reduce inequality between capital and labor, and if enough new tasks were targeting lower-skill workers, this could also reduce labor income inequality

What about a reduction in inequality because lower-skill workers benefit more? In the model here, the earnings of high-skill workers relative to low-skill workers is always pinned at λH/λU . So if new technologies reduce this ratio, they will reduce the gap between highskill and low-skill workers. But even this conclusion needs to be qualified. Acemoglu and Restrepo (2022) show that in more general settings, with multiple skill groups, there will be ripple effects whereby impacted demographic groups can then compete for tasks previously performed by other groups. In such a situation, an overall increase in labor productivity of both high-skill and low-skill workers in some tasks can lead to their displacement from these tasks, and then the ripple effects can, in principle, affect low-skill workers even more adversely than high-skill workers. While such adverse effects on low-skill workers are a general possibility in the framework of Acemoglu and Restrepo (2022), I am not aware of worked-out examples where an increase in the productivity of low-skill workers increases inequality. I now provide such an example.

“Suppose that labor productivity in z ∈ (I, I∗] increases due to advances in AI, and this also is more helpful for lower-skilled workers, so λH/λU declines. I now show that these advances could boost inequality. Suppose that after this increase in productivity, because σ < 1, the prices of tasks z ∈ (I, I∗] will decline and there will be less labor assigned to these tasks. If this effect is significant, all high-skill workers may be allocated away from these tasks, and the amount of labor demanded in these tasks may fall short of the supply of low-skill workers. In this case, the post-AI allocation may involve only low-skill workers performing tasks z ∈ (I, I∗], while both low-skill and high-skill workers perform tasks z ∈ (I∗, N ]. Then, regardless of how much λH/λU declines, the relative wage of skilled workers will be determined by the tasks that both types of workers are performing, which are now those above I∗, and thus will be equal to ω > λH/λU . Hence, inequality increases following the rise in the productivity of low-skill workers.”

“GDP therefore increases substantially more than average wages, and as a result, the capital share of national income increases by about 0.38 percentage points.”

“Nevertheless, there is no evidence that AI will reduce inequality between demographic groups, as some are forecasting. Rather, my analysis suggests that it may have a small positive effect on overall (between-group) inequality and reduce the real earnings of low-education women. It will also further widen the gap between capital and labor income.”

We then embed this framework in a dynamic economy in which capital accumulation is endogenous, and we characterize restrictions under which the model delivers balanced growth with automation and creation of new tasks, which we take to be a good approximation to economic growth in the United States and the United Kingdom over the last two centuries. The key restrictions are that there is exponential productivity growth from the creation of new tasks and that the two types of technological changes (automation and the creation of new tasks) advance at equal rates. A critical difference from our static model is that capital accumulation responds to permanent shifts in technology in order to keep the interest rate and hence the rental rate of capital constant. As a result, the dynamic effects of technology on factor prices depend on the response of capital accumulation as well. The response of capital ensures that the productivity gains from both automation and the introduction of new tasks fully accrue to labor (the relatively inelastic factor). Although the real wage in the long run increases because of this productivity effect, automation still reduces the labor share and employment.

For each of the forks in the road, the path that leads to a worse future is the one of least resistance and results in low productivity growth, higher income inequality, and higher industrial concentration. Getting to the good path of the fork will require hard work—smart policy interventions that help shape the future of technology and the economy. It is also important to appreciate a broader point about policy. Much of the discourse around AI regulation now takes place along a kind of hydraulic model: should we have more AI or less AI—or even ban AI. This discussion happens when AI is perceived as somewhat of a fixed thing, with a predetermined future. AI can come fast or slow. There can be more or less of it, but basically it is what it is. However, if policymakers understand that AI can develop in different directions, the discourse will be framed differently. How can policies encourage the types of AI that complement human labor instead of imitating and replacing it? What choices will encourage the development of AI that firms of all sizes can access, instead of just the largest ones? What kind of opensource ecosystem might that require, and how do policymakers support it? How should AI labs approach model development, and how should firms approach AI implementation? How does society get an AI that unleashes radical innovation, instead of marginal tweaks to existing goods, services, and systems?

To do so we first trace out the trajectories of U.S. commuting zones (CZs) in terms of the organization of local production, the education composition of the local population, income growth, and voting behavior in presidential elections from 1980 (the peak of manufacturing employment) to 2019 (before the onset of the Covid-19 pandemic).1 Over that time, manufacturing employment fell by roughly 6.5 million jobs, from 21% to 10% of U.S. workers (Pollard 2019). Accounting for aggregate growth in the labor force over time, the occupational displacement of production workers in the years following 1980 up through the 2008 financial crisis was of a magnitude similar to what LLMs might set in motion.

Beyond the similar potential magnitude of first order effects, the post-1980 manufacturing shock and its downstream effects serve as an apt comparison to LLMs’ projected effects because: 1) they marked a permanent shift as opposed to a cyclical movement; 2) they affected different geographic areas unevenly, and 3) they disproportionately affected a specific labor force segment—manufacturing production workers—that had been enjoying relative prosperity in the immediately preceding period.

Many of the effects of technological change in general and AI in particular are similar to those of globalization. Indeed, globalization can be viewed as a change in technology, that of trading with the rest of the world. In particular, trade of advanced countries with developing countries is “labor saving” (in the sense of Hicks): the demand for unskilled workers, or workers in general, decreases, at any given wage, implying that while the production possibilities curve moves out, and the utilities possibilities curve may move out, the new equilibrium entails workers being worse off, as in Figures 1 and 2. (In the absence of good risk markets, as we noted, everyone can be worse off, as in Figure 3). Thus, the issue of whether globalization is welfare enhancing comes back to the question addressed in this paper: is it possible to ensure, either through redistributive taxes or changes in institutions/rules, that workers are not made worse off. Again, there is a presumption that the gains to capital (or enterprises) could be taxed, to provide the requisite redistributions.13 As we discuss in greater detail below, one of the side effects of innovation and IPR is the creation of market power. Similarly, one of the consequences of globalization is to weaken the market power of workers. This is important because there is ample evidence that labor markets are far from perfectly competitive. The requisite compensation and/or offsetting changes in institutional rules to ensure that globalization represents a Pareto improvement may thus have to be all the greater.

Our analysis suggests that Wiener’s conjecture needs to be refined: different workers have different skills, and even if automated machines are like slave labor, they do not perfectly compete against all kinds of labor. Building on this intuition, we show that automation always reduces the real wages of worker types whose productivity schedule over tasks is sufficiently close to capital’s productivity schedule (if such worker types exist, but they may not).

The propagation matrix represents the full “ripple effects”—the impact of the displacement of one demographic group on others, as they leave the tasks they were previously performing and compete with other groups to be employed in other tasks. Such reallocations are the key channel via which direct productivity gains for a group may end up harming it at the end (as my example in the previous section illustrated). They are also the mechanism via which the displacement of a demographic group may end up being more damaging to another demographic group.

“Inequality is one of the main challenges posed by the proliferation of artificial intelligence (AI) and other forms of worker-replacing technological progress.”

“In a groundbreaking new study, Autor et al. (2024) link patent data to the creation of new job titles by census enumerators from 1940 onward to show how technological progress replaces some jobs while complementing the output of others and creating “new work.” They find that some patented innovations directly increase labor productivity, which in turn expands the set of tasks that workers do and leads to net employment growth. They call these changes augmentation innovations. Others, which they call automation innovations, generate employment declines. Moreover, Autor et al. (2024) show that during the 1940–1980 period, production occupations were highly exposed to automation innovations, while office clerks and other administrative-support workers were highly exposed to augmentation innovations. This difference is reflected in employment share declines for blue-collar production jobs and growth in clerical, office, and administrative-support jobs.”

On the second factor, the disruptions generated by AI‐related innovations depend on whether they are labor‐augmenting or labor‐saving, using the terminology of Hicks (1932), i.e. whether at a given wage, the innovations lead to more or less demand for labor. Some suggest that artificial intelligence will mainly assist humans in being more productive, and refer to such new technologies as intelligence assisting innovation, IA, rather than AI. Although we agree that most AI‐related innovations are likely to be complementary to at least some jobs – e.g. the ones applying AI to solve problems – we believe that taking a broader perspective, progress in AI is more likely to substitute for human labor, or even to replace workers outright, as we will assume in some of our formal models below.

AI-based productivity gains—measured either as growth of average output per worker or as total factor productivity growth—can come from a number of distinct channels：

Automation (or more precisely extensive-margin automation) involves AI models taking over and reducing costs in certain tasks. In the case of generative AI, various mid-level clerical functions, text summary, data classification, advanced pattern recognition, and computer vision tasks are among those that can be profitably automated.

Task complementarity can increase the productivity in tasks that are not fully automated and may even raise the marginal product of labor. For example, workers performing certain tasks may have better information or access to other complementary inputs. Alternately, AI may automate some subtasks, while at the same time enabling workers to specialize and raise their productivity in other aspects of their job.

Deepening of automation can take place, increasing the productivity of capital in tasks that have already been automated. For example, an already-automated IT security task may be performed more successfully by generative AI.

New tasks may be created thanks to AI and these tasks may impact the productivity of the whole production process.

Innovation also leads to large redistributions among others in the economy who are not directly involved in the process of innovation, for example workers who experience a sudden increase or decline in the demand for their labor. These redistributions can thus be viewed as externalities from innovation, and they are one of the main reasons why innovation raises concerns about inequality.

If – as many technologists predict – artificial intelligence directly replaces human labor, the demand for human labor will go down, and so will wages. More generally, innovations typically reduce demand for specific types of labor with specific human capital. For example, self‐driving cars will likely depress the wages of drivers, or radiology‐reading AI may lower the wages of traditional radiologists. Conversely, AI has certainly led to an increase in demand for computer scientists and has greatly increased their wages, in particular in sub‐fields that are directly related to AI. Since AI is a general purpose technology, there are reasons to believe that advances in AI will reverberate throughout many different sectors and lead to significant changes in wages throughout the economy in coming decades. Similar arguments can be made about the demand for and the value of different types of specific capital, as well as the demand for and prices of particular products.

Even though there are frequently losers, technological progress by definition shifts out the production possibilities frontier. This implies that the total dollar gain of the winners of progress exceeds the dollar loss of the losers lose

I find that commuting zones with a higher share of AI-adopting firms have experienced a stronger decline in the overall employment-to-population ratio and average wage during 2010-2021. The estimates suggest that a one standard deviation increase in AI exposure in the local labor market leads to 0.976 percentage points lower employment-to-population and 2.34% lower wage. Furthermore, the estimated effect implies that employment-topopulation in commuting zones at the 75th percentile of AI exposure declines by 1.25 percentage points more than in commuting zones at the 25th percentile of AI exposure, and average wage declines by 3% more.

The negative effect is heterogeneous. It is primarily borne by the manufacturing and low-skill services sectors, middle-skill workers, non-STEM occupations, and individuals at the two ends of the age distribution. The adverse impact is also more pronounced on men than women. These unequal effects of AI are similar as previous waves of labor market shocks, such as routine-biased technological change

However, higher exposure does not imply lower employment. Employment or hours worked can increase if AI complements human labor (Cazzaniga et al. (2024), Jiang et al. (2025)) or if the dispersion of task exposure to AI within an occupation is high

A decrease in educational attainment may not necessarily increase disparities, if LLMs can substitute for labor market skills and experience (Noy & Zhang 2023; Brynjolfsson et al. 2023). Autor (2024) interprets this possibility as a positive development that will “restore the middle class” by reducing the college/non-college wage gap. The argument is that “better” jobs will become newly available to the types of workers who fell behind due to the previously rising skill premium. In fact, whereas traditionally a college degree has been a prerequisite for many white-collar jobs, a recent analysis of trends in job postings reveals that employers are already beginning to remove a college degree as a job requirement, with about a fourfold increase from 2014 to 2023 in the annual rate of posted positions dropping the requirement (Fuller et al. 2024).

An alternative prediction is that the LLM shock will widen a different wage gap. Rather than the college/non-college framework, consider an LLM/non-LLM framework among holders of a college degree. Particularly if LLMs tend to substitute for certain occupations while complementing others, there may be a form of skill-biased technical change occurring among those with higher education. College graduates are not a homogeneously skilled demographic group, and there already exists significant variation in the returns to a college education by major (Altonji et al. 2012). The shift in labor demand induced by the LLM shock may well further widen the gap, especially in the MSAs with the greatest occupational exposure. As an illustrative example, the Wall Street Journal reports that in 2023, new job listings for workers within the information technology industry who have AI skills rose by 43%, whereas overall IT job listings declined by 31% (Rattner 2024). A similar trend has been documented in the academic literature (Acemoglu et al. 2022). If the effect of LLMs is to reduce the wage premium for a portion of college graduates, part of the rebuilding of a middle class as Autor (2024) envisions might develop not by bringing up workers from the bottom of the wage distribution, but by bringing down workers from the top. There may even be three groups to consider among college graduates: those directly involved in LLM work, who will likely be the biggest “winners”; those whose jobs are complemented by LLMs, who may become more productive and therefore experience wage gains, and those who are replaced by AI. It is this last group in particular that we predict is most likely to out-migrate in search of new labor market activities and more affordable living costs given potentially slower wage growth for their skill sets.

The fall of labor’s share of GDP in the United States and many other countries in recent decades is well documented but its causes remain uncertain.

In a letter to J. R. McCulloch written in June 1821, Ricardo wrote: “If machinery could do all the work that labour now does, there would be no demand for labour. Nobody would be entitled to consume any thing who was not a capitalist, and who could not buy or hire a machine” (Ricardo 1951–1973 8: 399–400).

In general, this expression has ambiguous sign, so automation can reduce wages. More specifically, there are two opposing effects (Acemoglu and Restrepo, 2018, 2019b): (a) automation always produces a positive effect on wages (and labor demand) because it increases productivity (or equivalently, reduces costs). This positive productivity effect is represented by the first term; (b) simultaneously, automation displaces workers from the tasks they used to perform. The negative displacement effect is represented by the second term. In the special case where R(K) is constant, it can be verified that automation increases wages. This is not the case, in general, when R(K) is increasing, as shown in Acemoglu and Restrepo (2018), because the displacement effect can be larger than the productivity gains.9 Overall, the impact of (extensive-margin) automation on the equilibrium wage is closely tied to its productivity effect, to which I next turn. Before doing so, I also note that the effects of task complementarities may be a little more complex than typically assumed. Even though an increase in γL(z) increases the marginal physical product of labor, the equilibrium wage is determined by the value of the marginal product of labor, which depends on the adjustment of task prices. As tasks produced by labor become more abundant/easier to perform, these task prices are reduced, and in the empirically relevant case where σ < 1 (as noted above), these task prices decline more than the increase in physical productivity. The equilibrium wage may still increase because of productivity gains, but the benefits to labor may be limited overall. For example, holding I and N constant, an increase in AL will leave wages constant when σ = sK where sK denotes the capital share in national income (Acemoglu and Restrepo, 2018). When σ < sK, higher AL can actually reduce real wages. Since sK ’ 0.4 currently in the US economy, Humlum’s estimates of σ mentioned above, of about 0.5, imply that task complementarities or labor-augmenting technological improvements will not raise wages much. Even when they increase wages, these technological shifts reduce the labor share, just like automation does.

We then embed this framework in a dynamic economy in which capital accumulation is endogenous, and we characterize restrictions under which the model delivers balanced growth with automation and creation of new tasks, which we take to be a good approximation to economic growth in the United States and the United Kingdom over the last two centuries. The key restrictions are that there is exponential productivity growth from the creation of new tasks and that the two types of technological changes (automation and the creation of new tasks) advance at equal rates. A critical difference from our static model is that capital accumulation responds to permanent shifts in technology in order to keep the interest rate and hence the rental rate of capital constant. As a result, the dynamic effects of technology on factor prices depend on the response of capital accumulation as well. The response of capital ensures that the productivity gains from both automation and the introduction of new tasks fully accrue to labor (the relatively inelastic factor). Although the real wage in the long run increases because of this productivity effect, automation still reduces the labor share and employment.

图i.1中美国的曲线表明了19102010年美国收入前10%人群的收入占国民收入的比重。这只不过是库兹涅茨针对19131948年这段时期建立的历史序列的延伸。收入前10%人群在20世纪前10年到20年代拥有了45%50%的国民收入，在20世纪40年代结束前该比例降到了30%35%。随后的19501970年，不平等程度一直稳定在该水平。到了20世纪80年代，我们看到不平等迅速增加，直到2000年美国的高收入阶层水平已回到占国民收入的45%50%。这一变化幅度令人印象深刻。我们自然会想知道这样一个趋势将持续多久。

美国收入前10%人群的收入占美国国民收入的比重从19101920年的45%50%下降到50年代的不足35%（这一下降被库兹涅茨记录在案）；之后该比重从70年代的不足35%上升到20002010年的45%50%。

图I.1 19102010年美国收入不平等

资料来源：piketty.pse.ens.fr/capital21c

1910年欧洲私人总财富的价值大约是67年的国民收入，在1950年价值约为23年的国民收入，在2010年价值为46年的国民收入。

图I.2 1870~2010年欧洲资本/收入比

资料来源：piketty.pse.ens.fr/capital21c

我将说明不平等的这一惊人增长很大程度上反映了高阶劳动收入的空前激增，大公司高级管理者在收入上将其他人远远甩在了身后。一个可能的解释是，这些高级管理者的技能和生产率较其他工人有了突飞猛进的增长。另一个解释是，这些高级管理者拥有制定自己薪酬的权力。这种权力在某些情况下没有限制，在更多的情况下与他们的个人生产率没有任何明确的联系，而在大型组织里个人生产率在任何情况下都难以有效评估。第二种解释在我看来更加合理，并且结果与证据更一致。这一现象最为显著的是在美国，在英国则程度轻一些，也许我们可以通过这两个国家过去一个世纪的社会和财政历史来解释它。在其他发达国家（如日本、德国、法国和其他欧洲大陆国家），这一趋势不是那么明显，但趋势的走向是相同的。后面我们会对这一现象进行全面的分析，预计这一现象在其他地方也会达到美国那样的程度——不幸的是，由于可用数据的限制，要做到全面分析并非如此简单。

首先从资本/收入比开始。收入是流量，它与某段时间内（一般为一年）生产和分配的产品数量相关。

资本是存量，它与某个时点上所拥有的财富总额相关，是此前所有年份获得或积累的财富总量。

衡量某个国家资本存量最自然而有效的方法是用这些存量除以每年的收入流量，从而得到资本/收入比，用希腊字母β表示。

例如，如果一个国家的资本存量总额等于6年国民收入之和，我们记为β=6（或者β=600%）。

在如今的发达国家，资本/收入比一般在56之间波动，而资本存量几乎完全由私人资本组成。2010年，英国、法国、德国、意大利、美国和日本的人均国民收入大约为3万3.5万欧元，而人均私人财富（除去债务）大约是15万20万欧元，是56年的国民收入。在欧洲和全世界，各国β值的差异也十分有意思。例如，日本和意大利的β值大于6，而美国和德国的β值则小于5。一些国家的公共财富略大于零，而其他一些国家则略小于零。我在之后几章会详细分析这个问题。此时，只要先了解这个基本数量级就可以了，这样会使概念更直观一些。[10]

事实上，2010年发达国家的人均国民收入为3万欧元（或2 500欧元每月），但显然并不意味着每个人都能获得这么多的收入。就像所有的平均数一样，这个人均收入数据也掩盖了极大的贫富差距。事实上，许多人的月收入低于2 500欧元，而有些人的月收入则是平均值的几十倍。收入差距主要由两个原因造成：一是劳动收入的不平等；二是资本收入的不平等，而这正是财富极端集中的后果。人均国民收入是指将既定总产出和国民收入在完全平均分配的条件下，分配给每个人的数量。[11]

同样，人均私人财富约为18万欧元（或相当于6年的国民收入），也并不意味着每个人都拥有这么多的财富。许多人拥有的资产要少得多，但是有些人的资产却高达几百万欧元乃至几千万欧元。许多人累积了很少量的财富——远低于他们一年的收入：或许有个几千欧元的银行存款，相当于几周或者几个月的工资。有些人的财产甚至为负：换句话说，他们拥有的资产要少于他们的负债。相比之下，有些人拥有大量的财富，可能是他们年收入的一二十倍甚至更多。一个国家的资本/收入比并不能反映国家内部收入不平等的情况，但是β值还是可以衡量一个社会总资本的重要性，因此分析这个比率是研究收入不平等的第一步。第二部分的主要目的是阐述不同国家的资本/收入比的差异、差异产生的原因以及它是如何随着时间演变的。

为了更好地理解目前世界上财富存在的具体形式，可以将发达国家的资本存量划分为大致相等的两个部分：家庭资本以及企业和政府使用的专业资本。总而言之，2010年，发达国家每位居民的年平均收入为3万欧元，拥有约为18万欧元的财富，其中9万欧元以房产的形式存在，另外9万欧元则是股票、债券、储蓄以及其他投资。[12]各国的情况各有区别，我将会在第二章讨论这个很有意思的问题。目前，记住资本可以分为大致相等的两个部分就可以了。

美国家庭正在变化，而且这种变化将导致不平等在下一代扩大。几乎可以肯定的是，有史以来第一次，美国的孩子们不会像其父母一样受到那么良好的教育，身体那么健康，或那么富有。

——朱恩·卡蓬和内奥米·卡恩（June Carbone and Naomi Cahn，2014）

衡量创新和技术变化的影响，就是看它们如何影响全要素生产率，全要素生产率是总体经济的一个平均数，等于表示经济总量的实际GDP除以资本与劳动投入的加权平均。谁也无法保证每一个社会成员都能平等共享经济进步的好处。本章将更加细致地分析收入分配的顶层与中层、底层之间非常不同的结果。去掉顶层人群的收入，余下99%人群所分配到的全部收入的增长率将低于全国总体收入的增长率。

顶层尤其是顶层1%人群的收入迅速上升可以由一系列推动最高收入上升的因素来解释，包括超级明星经济学、经理人薪酬激励的改变，以及房地产和股票市场的资本收益。收入分配底层90%人群的收入停滞则取决于一系列不同的原因，包括自动化带来的中等收入岗位减少、工会力量的削弱、最低工资购买力的下降、进口对制造业收缩的影响以及高技能和低技能移民的作用。

在本书中，“进步”都是由人均或时均实际GDP的提高速度来度量的。如果生活水平改善的速率使高收入群体比中等收入或低收入群体受益更多，那么，只讲这种平均数或均值就可能令人误解。当收入分配、财富分配或任何其他什么分配变得偏向于那些顶层群体时，数据序列的中位数就会比平均数增长得慢一些。这是过去40年里在美国真实发生的事情。在这一节中，我们研究三个不同的数据源，第一个基于税收记录，第二个基于美国人口普查局数据，第三个是考虑到税收和转移支付对低收入家庭税后收入再分配的影响，将税收和人口普查数据结合起来。

托马斯·皮凯蒂和伊曼纽尔·塞斯开创性地运用所得税数据研究顶层收入较之于顶层以下收入的演变。他们的数据追溯至1917年，即美国在1914年设立所得税短短几年后的数据。他们讨论了中低收入纳税群体的比例随年份更替和时代更迭而变化的问题，并提出了目前广为接受的解决方法，也就是运用国民收入账户的标准宏观经济数据估计总收入，然后基于纳税记录减去顶层收入，以获得低于顶层收入的那部分收入。

图18.1总结了皮凯蒂和塞斯关于1917—2013年近一个世纪的核心研究结论。显示的增长率在时间段上分为以1948年和1972年为间隔点的三个时间段，在收入获得者（income earners）上分为三组，即底层90%、顶层10%及覆盖全部收入获得者的平均收入。白色柱体显示底层90%群体税前收入（包括资本收益）的增长率，黑色柱体为顶层10%群体税前收入的增长率，灰色柱体为平均收入的增长率。三个时代中每一个都展示了迥异的结果。

图18.1　1917—2013年样本区间实际收入增长率

资料来源：根据Facundo Alvaredo、Anthony B.Atkinson、Thomas Piketty and Emmanuel Saez的世界顶层收入数据库计算，http://topincomes.g-mond.parisschoolofeconomics.eu/，2015年6月25日。平均收入指每一百分位数平均收入，包括资本收益。

1917—1948年，收入基本上趋于更加平等。底层90%收入群体的实际收入每年增长1.43%，是顶层10%收入群体增长率0.58%的两倍多，平均增长率为1.11%。这得益于大萧条、“二战”以及20世纪30年代至40年代的最低工资、立法鼓励工会和《退伍军人权利法案》等众多收入平等化计划，其中《退伍军人权利法案》让数以百万计的退伍军人进入大学，由工人阶层晋升为中产阶层。

1948—1972年的显著事实是底层90%、顶层10%和平均收入的增长率大致相同，而且每个组的实际收入增长都十分迅速。1948—1972年的实际收入平均增长率为2.58%，是1917—1948年平均增长率1.11%的两倍以上，1972—2013年平均增长率0.48%的5倍以上。1948年之后的25年是数以百万计高中毕业生的黄金时代，没有接受大学教育就可以在有工会的行业稳定工作，收入足以住上城市郊区有后院的房子，供一两辆车，过上大多数其他国家中等收入群体梦寐以求的生活。

但在20世纪70年代初之后，一切都发生了变化。收入分配底层90%与顶层10%群体之间的实际收入增长率形成巨大差距。2013年，底层90%的平均实际收入竟然比1972年还低。事实上，底层90%群体的实际收入峰值是2000年的37053美元，勉强高于1972年的35411美元，及至2013年这一平均数比2000年下降了15%，降至31652美元。同时，顶层10%群体的平均实际收入从1972年的161000美元增至2007年的324000美元，之后又下降至2013年的273000美元。

与收入不平等有关的第二个数据来源是人口普查局，它提供了1975年以来实际家庭收入平均数和中位数的数据。表18.1比较了来自人口普查局的1975—2013年收入增长率数据以及以1995年为界的两个子时段的收入增长率数据，并将皮凯蒂—塞斯自1975年以来的数据也置于同一张表中予以比较。如表18.1上部所示，来自人口普查局的1975—1995年实际收入平均数的增长率大大超过中位数的增长率，达0.61个百分点，1995—2013年超出0.33个百分点，两个时期合计的1975—2013年超出0.47个百分点。

表18.1　1975—2013年的实际收入增长率，以不同方式度量

资料来源：家庭收入中位数和平均数来自U.S.Census Bureau，Income and Poverty in the United States：2013，表A-1。百分位收入来自图18.1的数据。

表18.1下半部分展示了1975年以后皮凯蒂—塞斯数据中的增长率。各行分别显示顶层10%群体、底层90%群体和总体平均的每个应税单位的实际收入增长率，以及总体平均的收入增长率和底层90%的收入增长率之间的差异。全部三个时期所示的差异是0.70个百分点。有趣的是将人口普查局的平均收入增长与皮凯蒂和塞斯的平均收入增长进行比较；我们可以预期后者增长更快，因为这一概念包括资本收益，但在人口普查局的收入数据中资本收益被排除在外。尽管存在这种理论差异，在1975—2013年整个时期内，人口普查局的平均收入增长率年均0.77%，还是比皮凯蒂和塞斯的平均收入增长率年均0.60%略高一点。

最近有人在批评皮凯蒂—塞斯的数据以及人口普查局的数据时指出，它们只反映了来自市场的收入而忽视了税收和转移支付的影响。毫不奇怪，根据税收和转移支付调整后，包含了顶层收入组的平均收入与排除了顶层收入组的平均收入之间的增长率差异变小了。高收入个体所支付的税率大大高于大多数纳税人所支付的税率，实际上，在收入分配下半部分的大多数家庭很少或根本没有缴纳联邦所得税。社会保障、Medicare和雇主支付的医疗保险费是转移支付，有利于收入分配中层的群体，而食品券、所得税抵免和Medicaid转移支付主要针对收入分配底层家庭。

Technological change is never enough by itself to raise wages, however. Workers also need to get more bargaining power vis-à-vis employers, which they did in the second half of the nineteenth century. As industry expanded, firms competed for market share and for workers. Workers began to obtain higher wages through collective bargaining. This was the culmination of a long process that had started at the beginning of the century and reached fruition only in 1871, when trade unions became fully legal. This institutional transformation strengthened and in turn was supported by a broader push for political representation.

The beginnings of the computer revolution can be found on the ninth floor of MIT’s Tech Square building. In 1959‒1960, a group of often-unkempt young men coded there in assembly language into the early hours of the morning. They were driven by a vision, sometimes referred to as the “hacker ethic,” which foreshadowed what came to energize Silicon Valley entrepreneurs.

US median real wages (hourly compensation) grew at above 2.5 percent per year between 1949 and 1973. Then from 1980 onward, median wages all but stopped growing—increasing only 0.45 percent per year, even though the average productivity of workers continued to rise (with an annual average growth rate of over 1.5 percent from 1980 to the present).

This growth slowdown was far from equally shared. Workers with postgraduate degrees still enjoyed rapid growth, but men with a high school diploma or less saw their wages fall by about 0.45 percent, on average, every year between 1980 and 2018.

It was not just a widening gap between workers with postgraduate degrees and those with low levels of education. Every dimension of inequality skyrocketed from 1980 onward. For example, the share of the richest 1 percent of US households in national income rose from around 10 percent in 1980 to 19 percent in 2019. Wage and income inequality tells only part of the story. The United States used to pride itself for its “American dream,” which meant people from modest backgrounds rising in terms of income and children doing better than their parents. From the 1980s onward, this dream came under growing pressure. For children born in 1940, 90 percent of them earned more than their parents did, in inflation-adjusted terms. But for children born in 1984, the percentage was only 50 percent. The US public is fully aware of the bleak prospects for most workers. A recent survey by the Pew Research Center found that 68 percent of Americans think that today’s children will be financially worse off than their parents’ generation.

The distribution of income between capital and labor also changed significantly. Throughout most of the twentieth century, about 67‒70 percent of national income went to workers, and the rest went to capital (in the form of payments for machinery and profits). From the 1980s onward, things started getting much better for capital and much worse for workers. By 2019, labor’s share of national income had dropped to under 60 percent.

Labor’s share of national income has been on a protracted downward trend in most industrialized economies. In Germany, for example, it fell from close to 70 percent in the early 1980s to around 60 percent in 2015. At the same time, the income distribution became more skewed in favor of the very richest people. From 1980 to 2020, the share of the top 1 percent increased from about 10 percent to 13 percent in Germany, and from 7 percent to almost 13 percent in the UK. During the same period, inequality increased even in Nordic countries: the share of the top 1 percent rose from about 7 percent to 11 percent in Sweden and from 7 percent to 13 percent in Denmark.

Even at the best of times, the directions of technology and high wages are contested. Left to their own devices, many managers would try to reduce labor costs by limiting wage raises and also by prioritizing automation, which eliminates labor from some tasks and weakens the bargaining power of workers. These biases then influence the direction of innovation, pushing technology more toward automation.

Two other changes amplified the decline of labor and inequality. First, without countervailing powers from the labor movement, corporations and their managers developed a very different vision. Cutting labor costs became a priority, and sharing productivity gains with workers came to be viewed as akin to a failure of management. In addition to taking a harder line in wage negotiations, corporations shifted production toward nonunionized plants in the United States and increasingly abroad. Many firms introduced incentive pay, which rewarded managers and high performers, but at the expense of lower-skill workers. Outsourcing became fashionable as another cost-cutting strategy. Many low-skill functions, including cafeteria work, cleaning, and security, used to be performed by employees of large organizations such as General Motors or General Electric. These employees used to benefit from the overall wage increases that these companies’ workforces enjoyed. In the cost-cutting vision of the post-1980s, however, this practice was seen as a waste, so managers outsourced these functions to low-wage outside providers, severing another channel of wage growth for workers.

Second, it was not only companies choosing more automation from a given menu of technologies. With the new direction of the digital industry, the menu itself shifted powerfully toward greater automation and away from worker-friendly technologies. With a whole slew of digital tools enabling new ways of substituting machines and algorithms for labor, and little countervailing powers to oppose this move, many corporations embraced automation enthusiastically and turned their back on creating new tasks and opportunities for workers, especially those without a college degree. Consequently, although productivity (output per worker) continued to increase in the US economy, worker marginal productivity (how much that an additional hour of labor boosts production) did not keep up.

It bears repeating that shared prosperity was not destroyed by automation per se, but by an unbalanced technology portfolio prioritizing automation and ignoring the creation of new tasks for workers. Automation was also rapid in the decades following World War II but was counterbalanced by other technological changes that raised the demand for labor. Recent research finds that from 1980 onward, automation accelerated; more significantly, there were fewer new tasks and technologies that created opportunities for people. This change accounts for much of the deterioration of workers’ position in the economy. The labor share in manufacturing, where the acceleration of automation and the slowdown in the creation of new tasks has been most pronounced, declined from around 65 percent in the mid-1980s to about 46 percent in the late 2010s. Automation has also been a major booster of inequality because it concentrates on tasks typically performed by low- and middle-skill workers in factories and offices. Almost all the demographic groups that experienced real wage declines since 1980 are those that once specialized in tasks that have since been automated. Estimates from recent research suggest that automation accounts for as much as three-quarters of the overall increase in inequality between different demographic groups in the United States. The automotive industry is indicative of these trends. US car companies were some of the most dynamic employers in the country in the first eight decades of the twentieth century, and as we saw in Chapter 7, they were at the forefront of not just automation but also the introduction of new tasks and jobs for workers. Blue-collar work in the automotive industry was plentiful and well paid. Workers without college degrees and sometimes even without high school diplomas were hired and trained to operate new, sophisticated machinery, and they received quite attractive wages. The nature and availability of work in the automobile industry changed fundamentally in recent decades, however. Many of the production tasks in the body shop, such as painting, welding, and precision work, as well as a range of assembly jobs, have been automated using robots and specialized software. The wages of blue-collar workers in the industry have not increased much since 1980. Achieving the American dream through the automotive industry is much harder today than in the 1950s or 1960s. One can see the implications of this change in technology and organization of production in the hiring strategies of the industry. Since the 1980s, the US automotive giants stopped hiring and training low-education workers for complex production tasks and started accepting just higher-skilled applicants with formal qualifications, and only after a battery of aptitude and personality tests and interviews. This new human-resource strategy was enabled by the fact that there were many more applicants than available jobs and many of them had postsecondary education. The effects of automation technologies on the American dream are not confined to the automotive industry. Blue-collar jobs on other factory floors and clerical jobs in offices, which used to provide opportunities for upward mobility to people from disadvantaged backgrounds, have been the main target of automation by robots and software throughout the US economy. In the 1970s, 52 percent of US workers were employed in these “middle-class” occupations. By 2018, this number had fallen to 33 percent. Workers who once occupied these jobs were often pushed toward lower-paying positions, such as construction work, cleaning, or food preparation, and witnessed their real earnings plummet. As these jobs disappeared throughout the economy, so did many of the opportunities for workers with less than a postgraduate degree.

To be clear, I am not opposed to people making a lot of money. There’s a strong argument that it incentivizes economic growth under normal conditions. I am sympathetic to concerns about impeding innovation by killing the golden goose that generates it. But in a scenario where GDP growth is 10–20% a year and AI is rapidly taking over the economy, yet single individuals hold appreciable fractions of the GDP, innovation is not the thing to worry about. The thing to worry about is a level of wealth concentration that will break society.

The most famous example of extreme concentration of wealth in US history is the Gilded Age, and the wealthiest industrialist of the Gilded Age was John D. Rockefeller. Rockefeller’s wealth amounted to ~2% of the US GDP at the time.42

A similar fraction today would lead to a fortune of $600B, and the richest person in the world today (Elon Musk) already exceeds that, at roughly$700B. So we are already at historically unprecedented levels of wealth concentration, even before most of the economic impact of AI. I don’t think it is too much of a stretch (if we get a “country of geniuses”) to imagine AI companies, semiconductor companies, and perhaps downstream application companies generating ~$3T in revenue per year,43 being valued at ~$30T, and leading to personal fortunes well into the trillions. In that world, the debates we have about tax policy today simply won’t apply as we will be in a fundamentally different situation.

Related to this, the coupling of this economic concentration of wealth with the political system already concerns me. AI datacenters already represent a substantial fraction of US economic growth,44

and are thus strongly tying together the financial interests of large tech companies (which are increasingly focused on either AI or AI infrastructure) and the political interests of the government in a way that can produce perverse incentives. We already see this through the reluctance of tech companies to criticize the US government, and the government’s support for extreme anti-regulatory policies on AI.

“We should look to the history of our country here: even in the Gilded Age, industrialists such as Rockefeller and Carnegie felt a strong obligation to society at large, a feeling that society had contributed enormously to their success and they needed to give back. That spirit seems to be increasingly missing today, and I think it is a large part of the way out of this economic dilemma. Those who are at the forefront of AI’s economic boom should be willing to give away both their wealth and their power.”