Research Library: Studies, Resources, and Theories

Summary

This chapter provides the basic economic language needed to discuss automation without reducing the subject to a simple count of jobs gained or lost. Acemoglu and Restrepo begin from the idea that production is made up of many separate tasks. Some tasks are performed by workers, some by machines, and some through combinations of both. AI and automation matter because they change the boundary between these categories. When a machine becomes capable of performing a task previously assigned to labor, the authors call the result a displacement effect. The immediate consequence is lower demand for workers in that task and a tendency for labor’s share of the resulting income to fall.

Automation may also make production cheaper and more efficient. Lower costs can reduce prices, raise output, increase investment, and create demand in related activities. The authors call this the productivity effect. Capital accumulation can reinforce it: once automated production becomes profitable, firms buy more equipment and expand capacity. These forces may offset some employment losses. However, the paper emphasizes that higher productivity does not automatically restore workers’ bargaining position or their share of national income. Output per worker may rise faster than wages, allowing owners of capital to capture a larger fraction of the gains.

The strongest counterweight to displacement is the creation of new tasks in which people have a comparative advantage. New technologies can generate occupations, responsibilities, and services that did not previously exist. This reinstatement effect moves part of production back toward labor and can raise both employment and labor’s income share. The historical success of industrial economies therefore depended not only on automating old work but also on continually inventing new work for people.

The authors also stress that transitions are slowed by real-world frictions. Workers may not possess the skills required by new tasks; firms may automate because machines are subsidized or fashionable even when the productivity improvement is modest; and education, institutions, and labor markets may adjust too slowly. Their position is neither that automation inevitably creates mass unemployment nor that markets automatically make everyone better off. The outcome depends on the balance among displacement, productivity, capital accumulation, and new-task creation. For the Capitalism without Labor, this study supplies the starting framework: rising output can coexist with weak wages, and employment can recover while ownership income becomes increasingly dominant.

Why it matters to Levels 1–9

This is the theoretical starting point for the Capitalism without Labor. It explains early and middle transition levels and identifies the central long-run issue: output can rise while labor income weakens.

Limit or caution

This is a framework, not a timetable. It does not predict when particular occupations will disappear or fully describe a Level 9 economy.

Summary

This study develops the task framework into a historical explanation of why technological progress sometimes strengthens labor and sometimes weakens it. The authors distinguish two opposing changes in the content of production. Automation transfers tasks from workers to capital, creating a displacement effect. The introduction of genuinely new tasks gives workers new areas of comparative advantage, creating a reinstatement effect. Productivity growth can support labor demand, but the allocation of tasks determines who performs the expanding production and who receives the income.

The central contribution is to show that these forces can be inferred from industry data rather than treated only as abstract theory. If productivity rises while labor’s share falls, automation is likely shifting the task boundary against workers. If labor’s share and employment rise because new activities require people, reinstatement is occurring. The authors use this logic to interpret changes in the United States after World War II. In earlier decades, automation was accompanied by substantial creation of new occupations and tasks. New work in professional, technical, managerial, clerical, and service activities helped absorb workers released from older production methods.

The balance became less favorable in more recent decades. The authors conclude that the slower growth of employment can be explained by a faster displacement effect, particularly in manufacturing, a weaker reinstatement effect, and slower productivity growth than in the earlier postwar period. This combination is especially damaging: workers lose tasks rapidly, too few valuable new tasks appear, and the productivity gains are not large enough to create strong demand elsewhere. The paper therefore rejects the comforting assumption that technological change always creates enough new work simply because it did so in the past.

For our Levels 1–9 framework, the study suggests that the crucial variable is not merely the percentage of existing jobs that can be automated. We must also track the rate at which new human tasks are created, the quality and compensation of those tasks, and how long they remain protected from the next wave of automation. At lower levels, reinstatement may continue to support employment. At higher levels, AI may begin to automate the new tasks soon after they emerge, weakening the historical adjustment mechanism. The authors do not claim that this outcome is inevitable, but their framework shows exactly what must be measured: the economy needs a continuing stream of productive roles for humans if labor income is to remain central.

Why it matters to Levels 1–9

This directly supports our need to track not only jobs eliminated, but also the rate and quality of new human occupations created at every level.

Limit or caution

New tasks may be temporary if later generations of AI and robotics can automate them too. The historical record may therefore understate Level 9 risks.

Summary

This paper asks a narrower and more empirical question than the task-based theory papers: what actually happened in American communities that were more exposed to industrial robots? The authors study the spread of robots from 1990 to 2007, concentrating on machines used in manufacturing activities such as welding, painting, assembly, material handling, and packaging. Because robot adoption is not evenly distributed across industries, local labor markets with different industrial compositions received different levels of exposure.

The analysis is organized around U.S. commuting zones, which approximate local labor markets. The researchers construct a measure of predicted robot exposure using the industries present in each area and the pace at which those industries adopted robots. They then compare changes in employment and wages across more- and less-exposed areas while attempting to separate robot effects from other forces such as Chinese import competition, offshoring, ordinary computer investment, and the decline of routine work.

The main finding is negative for workers in exposed communities. The authors estimate that one additional robot per thousand workers reduced the employment-to-population ratio by roughly 0.18 to 0.34 percentage points and average wages by about 0.25 to 0.5 percent. The effects were particularly relevant to manufacturing and routine production workers, but they also spread into surrounding local services because displaced workers and weaker wages reduced local purchasing power. The paper’s estimates imply that a robot can eliminate several jobs in the affected local market, even though the machine may also raise productivity and profitability for the adopting firm.

This study is important because it demonstrates that adjustment is not painless or geographically neutral. National employment statistics may conceal severe losses in factory communities, and workers cannot always move quickly into expanding sectors or regions. It also illustrates the difference between a technology’s technical success and its social consequences. The study does not prove that every future robot will destroy jobs, and it covers an earlier generation of specialized industrial machines rather than flexible humanoids or general AI. Still, it provides a concrete warning for the Ford example: when physical automation reduces the need for production labor, the surrounding community can lose wages and demand long before the broader economy develops replacement opportunities.

Why it matters to Levels 1–9

It provides concrete evidence for the first physical-robot transition in manufacturing and is particularly relevant to the Ford example.

Limit or caution

The study covers an earlier generation of industrial robots and local effects. It does not establish the economy-wide result of future general-purpose robotics.

Summary

David Autor’s essay is a history of how economists have changed their understanding of technology and work. He organizes several decades of research into four broad intellectual periods. The first was the “education race,” when computers appeared to reward educated workers and the main policy response was to expand schooling. The second was the task-polarization model, which explained why computers displaced routine middle-skill work while complementing abstract professional work and leaving many manual service jobs intact. This helped explain the growth of high-wage and low-wage occupations alongside the decline of many middle-income jobs.

The third period focused on the race between automation and reinstatement. Here, the important question became whether technology was merely replacing workers in existing tasks or also creating new tasks that restored labor demand. This framework produced a more cautious form of optimism: technology could still generate shared prosperity, but only when institutions, investment, education, and innovation created sufficiently valuable roles for people.

Autor describes the present AI period as one of “vast uncertainty.” Modern AI reaches into nonroutine cognitive work that earlier automation theories often treated as protected. Writing, coding, diagnosis, analysis, design, and professional judgment may be partly automated or transformed. At the same time, technical exposure does not reveal how firms will reorganize jobs, how rapidly adoption will occur, whether demand will expand, or which new tasks will emerge. Historical analogies are informative but not decisive because the range and adaptability of AI may differ from earlier machines.

The essay’s most important position is that the future should not be treated as a technological fate waiting to be predicted. Societies choose which technologies to develop, how firms deploy them, how workers share in productivity gains, and which institutions support adaptation. Policies affecting education, labor standards, competition, research incentives, and worker voice can influence whether AI complements expertise or primarily substitutes for labor. For the Capitalism without Labor, this study gives us a disciplined attitude: we should neither assume that the Industrial Revolution guarantees a benign outcome nor assume that every exposed job disappears. The transition is uncertain because technology and institutions interact, and that uncertainty makes scenario planning more appropriate than one rigid forecast.

Why it matters to Levels 1–9

This resource helps the Atlas remain open to several transition pathways and avoid presenting Level 9 as a predetermined social system.

Limit or caution

It is a conceptual review rather than a quantitative forecast or policy blueprint.

Summary

Autor and Salomons investigate an apparent contradiction. Automation clearly replaces workers inside particular activities, yet advanced economies have not experienced a permanent collapse in total employment. The authors separate what happens inside an automating industry from what happens across the rest of the economy. They analyze four decades of harmonized industry data from multiple countries, using common movements in industry productivity as a measure of technological progress.

Inside the industry where automation originates, the direct effect is labor-displacing. Firms produce more with fewer workers, and labor’s share of the industry’s value added falls. But several indirect channels can create employment elsewhere. Lower costs may reduce prices and expand demand for the industry’s product. Automating suppliers can lower input costs for customer industries, allowing those customers to expand. Consumers who benefit from lower prices may spend their savings on other goods and services. Employment can also shift toward sectors where demand grows faster.

The authors find that these indirect effects are strong enough, in their historical data, to reverse the direct employment losses at the aggregate level. In other words, technological progress can reduce jobs in the originating industry while supporting total employment through expansion elsewhere. This helps explain how agricultural and manufacturing employment could shrink dramatically without causing permanent mass unemployment. However, the same recovery does not occur for labor’s share of income. The income lost by workers in automating industries is not fully restored in other sectors. Aggregate employment may remain resilient while a larger portion of national income flows to capital.

This distinction is central to our project. Counting jobs alone can produce a falsely reassuring conclusion. A future economy may still employ many people, but in lower-paid services, shorter-hour positions, or roles with weak bargaining power, while owners of automated capital receive most productivity gains. The study also warns against looking at Ford in isolation. Ford may reduce employment, lower vehicle costs, and stimulate other industries; those indirect effects could create work. Yet the distribution of income can still move away from labor. The paper is historical rather than a forecast of Level 9, and its aggregate employment result may not hold when AI and robotics automate most sectors simultaneously. Its durable lesson is that employment quantity and income distribution are separate questions.

Why it matters to Levels 1–9

This is essential for our sector method: a sector cannot be evaluated in isolation. It also warns that “jobs recovered elsewhere” does not mean workers retain the same share of prosperity.

Limit or caution

The study measures automation through common industry productivity movements, not direct observation of every technology. Future AI may behave differently.

Summary

This study provides a detailed baseline of how widely American firms were using advanced technologies before the generative-AI boom. It draws on a special module of the U.S. Census Bureau’s 2019 Annual Business Survey, covering more than 300,000 firms across the economy. The survey asks about five categories: artificial intelligence, robotics, dedicated equipment, specialized software, and cloud computing. This matters because public discussion often assumes that a technology is already widespread merely because it is technically impressive.

The authors find that adoption was uneven and, for AI and robotics, still relatively uncommon. Cloud computing and specialized software were more widely used, while robotics remained concentrated in particular manufacturing and logistics applications. Large firms were much more likely to adopt than small firms, and younger firms were often more technology-intensive. Because large adopters employ many people, the share of workers employed at firms using a technology can be much larger than the share of firms adopting it.

The paper also examines how technology use is associated with firm characteristics and workforce composition. Adopting firms tend to be more productive, employ more highly educated and specialized workers, and invest in complementary organizational assets. Some technologies are associated with lower labor shares or changing skill requirements. However, the authors are careful about causality: productive firms may be more capable of adopting technology, so the technology itself may not be the sole reason for their performance. At this early stage, employment effects are limited or ambiguous rather than evidence of immediate mass displacement.

For the Levels 1–9 transition, the study demonstrates why adoption speed must be separated from capability. Even if a robot or AI system can perform a task, firms need capital, data, expertise, compatible processes, and confidence that the investment will pay. Large companies may move first, creating competitive pressure that later forces smaller firms to adopt or exit. The survey therefore represents Level 1 and early Level 2: islands of advanced automation within an economy still organized around human work. It gives us a starting point for measuring diffusion and reminds us that the transition will likely be led by a relatively small number of large, capable firms before it becomes economy-wide.

Why it matters to Levels 1–9

This provides a factual Level 1 baseline and shows why transition can begin through uneven adoption by a relatively small number of powerful firms.

Limit or caution

The data capture early adoption and reported effects, not mature AI-agent or general-purpose robot systems.

Summary

This paper resolves a puzzle that often confuses discussions of automation: a firm that buys robots may expand employment even though robots reduce employment in the industry as a whole. The authors construct a dataset of robot purchases by French manufacturing firms between 2010 and 2015. Only 598 of the 55,390 firms in the sample adopted robots, but these adopters were unusually important, accounting for about one-fifth of manufacturing employment and value added.

Robot-adopting firms became more productive and increased value added. Their labor share fell, and production workers became a smaller share of their workforce, which is consistent with machines replacing some production tasks. Yet the adopters also expanded total employment. At first glance, this might suggest that robots create jobs. The competitive analysis changes the interpretation. Automation lowers an adopter’s relative costs and improves its ability to gain market share. The firm expands partly by taking sales from competitors that have not automated.

When the contraction of competing firms is included, the industry-wide employment effect is negative. The paper therefore distinguishes between a private business result and a social result. For the adopting firm, robots can support growth, profits, and even additional hiring in complementary activities. For the industry, the same technology can reduce the total number of workers required to produce a given amount of output. The decline in the labor share is also stronger at the industry level because adopters are large and become larger.

This evidence is highly relevant to a future Ford transition. Ford might automate, lower costs, gain market share, and report that its remaining employment is stable or growing in engineering and software. That would not mean automation had no displacement effect. Less automated competitors and suppliers might shrink, and the entire automotive sector could employ fewer people. The study also shows why market competition accelerates adoption: once one large firm becomes significantly cheaper, others must automate to survive. Its limitation is that it covers specialized industrial robots over a short period in one country. Nevertheless, it offers a powerful mechanism for the middle transition levels: automation spreads not only because every investment is intrinsically superior, but because early adopters change the competitive environment for everyone else.

Why it matters to Levels 1–9

This is a key warning for Ford-style case studies. Ford could prosper and even retain some workers while the automotive ecosystem loses substantially more employment.

Limit or caution

The sample is French manufacturing from 2010–2015; results may not generalize mechanically to every service sector or future AI system.

Summary

The ILO study is designed to measure which occupations could be affected by generative AI without equating exposure with unemployment. It updates an earlier global index to reflect rapid improvements in language and multimodal systems. The researchers begin with detailed task descriptions—nearly 30,000 tasks in the occupational classification used for the project—and combine worker judgments, expert review, and AI-assisted assessment. Occupations are then placed into graded exposure categories rather than labeled simply “safe” or “automatable.”

The report finds that clerical work remains the most exposed because many duties involve producing, processing, organizing, and communicating information in standardized formats. Exposure is also rising in professional, technical, media, finance, and administrative occupations as models become capable of handling more complex text, image, and analytical tasks. Roughly one in four workers worldwide is employed in an occupation with some degree of exposure, but only a smaller fraction falls into the highest category.

The ILO’s principal conclusion is that transformation is more likely than complete job replacement for most occupations at current capability levels. A job usually combines tasks with different technical, legal, interpersonal, and physical requirements. AI may automate document drafting, scheduling, or analysis while leaving responsibility, negotiation, fieldwork, care, or human interaction to workers. Employers may redesign the job, reduce staffing, raise output, or change skill requirements; the exposure index does not determine which response will occur.

The report also highlights differences among countries and demographic groups. High-income economies generally contain more information-intensive occupations and therefore show greater exposure. Women are often more exposed because they are overrepresented in clerical and administrative roles. At the same time, countries with weak digital infrastructure may have high theoretical exposure but slow actual adoption. For our transition model, the ILO index is a useful map of the cognitive frontier at Levels 2–4. Its greatest value is methodological discipline: it tells us where capabilities overlap with tasks, not how many people will certainly lose jobs. It also warns that the same technology can produce different outcomes depending on infrastructure, institutions, training, and worker participation.

Why it matters to Levels 1–9

This is one of the best resources for constructing a Level 1–4 occupational map, particularly for office and cognitive work.

Limit or caution

It measures capability exposure, not the speed of adoption, profitability, regulation, social acceptance, or physical robotics.

Summary

The OECD report examines AI as a workplace institution, not merely as a job-counting technology. It combines labor-market statistics, reviews of existing research, and surveys of more than 2,000 employers and 5,300 workers in manufacturing and finance across seven OECD countries. These sectors were chosen because they had meaningful experience with technologies such as computer vision, machine learning, and natural-language processing before the broad diffusion of generative AI.

At the time of the study, the OECD found no clear evidence that AI adoption had reduced economy-wide labor demand. Most surveyed firms reported no change in employment, and the difference between firms reporting increases and decreases was not statistically decisive. Early applications often assisted skilled workers rather than eliminating their positions. The report therefore rejects the claim that technical progress had already produced mass AI unemployment.

However, employment totals are only part of the story. AI can change work intensity, autonomy, monitoring, scheduling, evaluation, and the distribution of decision-making. Workers often reported that AI helped them perform tasks and improved some aspects of job quality, but many also feared job loss, wage pressure, unfair surveillance, and loss of control. Benefits were more likely when workers received training and had a voice in implementation. Risks increased when systems were imposed without transparency or used primarily for monitoring and cost reduction.

The OECD’s answer is institutional: AI’s labor-market effects are not fixed by the software alone. Management practices, regulation, collective bargaining, social dialogue, education, and competition policy influence whether AI complements workers or weakens them. This makes the report particularly important to the Capitalism without Labor. At early levels, the policy challenge is not yet to distribute the output of a fully automated economy; it is to shape how firms introduce AI, who participates in decisions, and who receives productivity gains. The report is cautious because its evidence predates much of the generative-AI wave and focuses on early adopters. Still, it establishes a key principle for later chapters: the quality, power, and compensation attached to remaining work matter as much as the number of jobs.

Why it matters to Levels 1–9

This broadens our transition model beyond unemployment and helps define the institutional changes required before labor becomes economically secondary.

Limit or caution

The report predates the most capable agentic systems and focuses on OECD labor markets rather than a Level 9 end state.

Summary

This study was one of the first systematic attempts to estimate how large language models could affect the U.S. labor market. The title plays on two meanings of GPT: generative pretrained transformer and general-purpose technology. The authors use detailed occupational task data and a new scoring rubric to ask whether current language models—or software systems built around them—could reduce the time required to complete tasks while maintaining acceptable quality.

The analysis distinguishes direct model capability from the broader impact of LLM-powered applications. A stand-alone chatbot may assist with drafting or classification, while integrated software can connect the model to databases, workflows, and specialized tools. Human evaluators and GPT-4 were both used to classify task exposure. The authors estimate that approximately 80 percent of U.S. workers could have at least 10 percent of their tasks affected, and about 19 percent could have at least half of their tasks affected.

Unlike many earlier waves of automation, exposure is substantial in higher-income, educated, and information-intensive occupations. Writing, programming, legal analysis, finance, administration, and other cognitive activities appear prominently. Physical and outdoor work is generally less exposed because the study concerns language models rather than robotics. This reversal is important: AI does not simply continue the old pattern of replacing routine factory labor while protecting professionals.

The authors repeatedly caution that exposure is not a forecast of job loss. A time-saving tool may increase output, lower prices, improve quality, or allow workers to concentrate on other duties. Jobs consist of bundles of tasks, and employers decide whether to reduce staffing, redesign work, or expand services. The estimates also depend on model capability and software integration, both of which change rapidly. For our Levels 1–9 framework, the paper maps an early cognitive automation frontier and demonstrates why software AI may spread faster than physical robots. It supports the expectation that middle transition levels will affect office and professional work alongside manufacturing. Its broader claim—that LLMs have characteristics of a general-purpose technology—suggests that their largest effects may come from organizational redesign and complementary innovations rather than from isolated chatbot use.

Why it matters to Levels 1–9

It is a useful early map of the cognitive side of the transition and a reminder that automation will not move simply from low-skilled to high-skilled work.

Limit or caution

Capability exposure is not adoption, job loss, or a prediction of future model ceilings. The methodology was based on early GPT-4-era capabilities.

Summary

The World Economic Forum report is a survey of employer expectations rather than an independent prediction generated from a single economic model. It gathers responses from more than 1,000 large employers representing over 14 million workers, 22 industry groups, and 55 economies. Employers are asked how technological, demographic, environmental, geopolitical, and economic trends are likely to change their workforce between 2025 and 2030.

When the survey responses are combined with International Labour Organization employment data, the report estimates that structural change could create about 170 million jobs and displace about 92 million by 2030, producing a net gain of roughly 78 million. These figures include more than AI: digital access, the green transition, population aging, geoeconomic fragmentation, and economic uncertainty all contribute. AI and information-processing technologies are expected to create and displace large numbers simultaneously, while robotics and autonomous systems are expected to be a significant net displacer.

The expected occupational pattern is mixed. Clerical and routine administrative roles are among the largest declining categories. Technology occupations, care work, education, construction, delivery, agriculture, and green-transition roles are expected to grow in many regions. Employers anticipate that about 39 percent of workers’ core skills will change by 2030 and identify skill gaps as the leading barrier to business transformation. Training, redeployment, hiring for new skills, and automation are all part of their planned responses.

The report’s value for our project is that it shows how major employers currently imagine the near-term transition. They do not expect a sudden move to Level 9; they expect simultaneous creation, destruction, and redesign of jobs. It also provides sector and regional detail that can guide later chapters. Its limitation is equally important: employer plans can be wrong, may reflect corporate interests, and cover only a five-year horizon. A positive net job estimate through 2030 does not answer what happens when AI and robotics become capable of automating the newly created work. We should therefore use this report as a Level 1–3 expectations map, not as proof that long-run technological unemployment is impossible.

Why it matters to Levels 1–9

It provides a near-term expectations baseline against which actual Level 2–3 outcomes can later be compared.

Limit or caution

Employer expectations are not forecasts with guaranteed accuracy. The report ends in 2030 and cannot answer whether net job creation continues at higher automation levels.

Summary

This paper explains why a transformative technology can appear disappointing in official productivity statistics during its early years. General-purpose technologies such as electricity, computers, and AI do not create their full value when a firm simply purchases equipment or software. Organizations must invent complementary processes, products, business models, data systems, training programs, and management structures. These investments are often intangible and are poorly captured in conventional accounting.

During the implementation period, firms spend money and employee time on experimentation and reorganization. These activities may reduce current measured output even though they are building valuable future capabilities. Accounting systems often record the costs as ordinary expenses rather than as investment in a new productive asset. As a result, measured productivity can slow or decline while the underlying technological foundation is being constructed.

Once the complementary assets are in place, the technology begins to generate large benefits. Firms harvest the value of earlier redesign, and measured output may rise rapidly. Because the earlier intangible investment was understated, later productivity growth can be overstated relative to the measured capital stock. This produces the J-shaped pattern: an initial dip followed by a later acceleration. The authors develop a formal model and relate it to historical evidence involving research and development, software, and computer hardware.

For the Capitalism without Labor, the J-Curve helps explain why the transition may be slow at first and then accelerate. A company like Ford must redesign vehicles, factories, supply chains, quality systems, software, and management before advanced robots can replace most labor. Early costs may be large and the productivity benefits modest. Once the new production architecture is proven, replication across plants may become much faster. The theory also cautions against declaring AI either a failure or an immediate revolution based on short-term productivity data. It does not predict how gains will be distributed or whether employment will fall. Its focus is timing and measurement: organizational transformation is part of the technology, and the economic payoff may arrive only after years of hidden investment.

Why it matters to Levels 1–9

This is directly relevant to the Level 1–5 transition. AI and robotics may appear economically modest before organizational redesign produces a rapid acceleration.

Limit or caution

The theory explains delayed gains but does not guarantee that every technology eventually produces large gains or that the gains are broadly shared.

Summary

This field study examines generative AI inside an actual workplace rather than estimating exposure from task descriptions. The researchers follow the staggered introduction of an AI conversational assistant to 5,179 customer-support agents. The system listens to customer interactions and suggests responses, information, and guidance based partly on patterns learned from successful previous conversations.

Access to the tool increased productivity—measured as customer issues resolved per hour—by about 14 percent on average. The effects were highly unequal across workers. Novice and lower-skilled agents improved by roughly 34 percent, while experienced and highly skilled workers saw little benefit. The evidence suggests that the AI captured and transmitted practices associated with top performers, allowing newer workers to move down the learning curve faster.

The study also reports improvements in customer sentiment and some evidence of higher employee retention and worker learning. These findings matter because they show that AI can change the distribution of expertise. Knowledge that previously required months of experience can be embedded in a system and delivered at the moment it is needed. That can make less-experienced workers more productive and narrow performance differences inside the occupation.

The immediate result is augmentation, not full replacement. Agents still conduct the conversation, interpret the situation, and take responsibility, while the system supplies recommendations. Yet the same productivity improvement can eventually affect staffing. If a fixed volume of customer inquiries can be handled with fewer agents, firms may reduce hiring even without layoffs. Alternatively, lower service costs may expand demand and improve service quality. The study does not settle which long-run effect dominates.

For our transition framework, this is a clear Level 2 example: AI assists workers and spreads high-level practices before jobs are fully automated. It also raises a deeper question. If AI makes novice workers nearly as effective as experts, the market value of accumulated experience may fall, even while overall productivity rises. The study is limited to one company and one structured form of customer support, but it provides strong evidence that generative AI can rapidly reorganize skills, training, and productivity inside an occupation.

Why it matters to Levels 1–9

It is a concrete example of early transition: AI can first compress skill differences and increase output per worker, which may later reduce hiring needs even without immediate layoffs.

Limit or caution

The setting is one occupation and one AI deployment. Long-term staffing, wage, and competitive effects were outside the study’s main window.

Summary

This IMF working paper builds macroeconomic models in which robots are a special form of capital capable of substituting for human labor. Its provocative title does not mean that automation lowers total output. On the contrary, the models generally produce more goods and services. The concern is that the transition can distribute those gains very unevenly and can make large groups of workers worse off for a long time.

When robot productivity improves, firms substitute machine services for human labor. The supply of effective labor-like services rises, reducing the scarcity value of competing workers. Real wages can fall, while returns to capital and the income of robot owners rise. Higher savings and investment expand the robot stock, which increases production but can reinforce the pressure on wages. Depending on the model’s assumptions, wages may eventually recover as capital accumulation and output become enormous, but the recovery can take generations.

The paper examines different possibilities, including distinctions between skilled and unskilled labor and cases in which robots substitute more strongly for one group. Workers who own little capital are vulnerable because their principal asset is their labor. Skilled savers and existing capital owners benefit earlier. Even future generations may be worse off in some scenarios if inequality, saving behavior, and substitution create what the authors describe as highly unfavorable transition dynamics.

The main conclusion is therefore not “stop technology.” It is that aggregate growth is an insufficient measure of welfare. A society can become much richer in productive capacity while workers experience falling wages, weak bargaining power, and reduced consumption. Policies that broaden ownership, redistribute gains, support consumption, or influence the pace and form of automation may be necessary.

This paper is one of the closest formal analogues to our Ford thought experiment. Ford’s robot capital can increase output and profits while displaced employees lose the wage income needed to participate in the market. At Levels 5–8, that pattern could spread across sectors. The models are stylized and do not predict a precise future, but they establish the central Post-Labor Capitalism problem: if labor becomes abundant and machine capital remains concentrated, economic abundance does not automatically translate into broad prosperity.

Why it matters to Levels 1–9

This is one of the closest existing studies to our original Ford thought experiment and the difficult middle between Level 1 and Level 9.

Limit or caution

Results depend on model assumptions about substitutability, saving, skills, and ownership. The provocative title should not be read as an empirical certainty.

Summary

Korinek and Stiglitz examine the distributional consequences of worker-replacing technology and challenge a common economic shortcut. Standard theory often says that technological progress creates enough additional output to compensate anyone who loses. The authors point out that the possibility of compensation is not the same as actual compensation. Markets do not automatically transfer gains from technology owners to displaced workers.

The paper develops a taxonomy of circumstances under which AI can improve or reduce individual welfare. Outcomes depend on whether AI substitutes for or complements different kinds of labor, who owns the productive capital, how prices change, and whether workers can move into new tasks. Even when total national income rises, workers may face lower wages or unemployment. If they own little capital, they may not share meaningfully in the gains.

The authors also emphasize market imperfections. Innovation decisions are driven by private profit, which may not equal social value. Firms may prefer technologies that save labor costs even when other innovations would create greater overall productivity or social benefit. Monopoly power, incomplete insurance, political influence, and weak redistribution can intensify the losses. The distribution of ownership and the direction of technological research are therefore part of the economic mechanism, not afterthoughts.

Policy can affect both the transition and the final distribution. The paper considers redistribution, taxation, social insurance, education, and ways to steer innovation toward complementing people. It does not present one final institutional design, but it rejects the idea that growth alone guarantees a Pareto improvement in which everyone benefits. People who lose their labor income may require direct claims on the gains or access to new productive assets.

For the Capitalism without Labor, this source helps define the difference between production and distribution. Level 9 may solve the technical problem of producing goods with little human labor while worsening the economic position of people who do not own the systems. The paper supports our decision to study ownership, purchasing power, and policy alongside automation. It also leaves open the entrepreneurial question: how can society preserve incentives for invention while preventing the gains from becoming so concentrated that most citizens are economically dependent?

Why it matters to Levels 1–9

This study provides the intellectual bridge from automation economics to Post-Labor Capitalism: production, ownership, and distribution must be analyzed together.

Limit or caution

The paper maps possibilities rather than specifying one operational constitution for a post-labor system.

Summary

This IMF paper addresses competing claims about AI and inequality. One view predicts that AI will widen disparities, as previous automation often did. Another argues that generative AI may reduce wage inequality because it reaches high-income white-collar occupations rather than concentrating on routine, lower-paid work. The authors combine household microdata with a calibrated task-based macroeconomic model to separate the channels behind these claims.

The first channel is labor displacement. High-income workers are often employed in occupations with substantial AI exposure. If AI replaces their tasks, wage gaps could narrow because higher earners lose more. The second channel is complementarity. Many high-income professionals may use AI to become more productive rather than being replaced. If AI raises the productivity of their judgment, management, or specialized knowledge, their wages may remain strong or increase.

The third channel is capital ownership. High-income households are more likely to own stocks, businesses, and other assets. When firms adopt AI to save expensive labor, the returns to automated capital can rise. The same people whose occupations are exposed may therefore benefit through investment income. Once firms’ adoption choices are included, the model finds that wealth inequality is likely to increase substantially even in scenarios where wage inequality falls.

This distinction is especially important for our project. A society might observe narrowing wage differences and mistakenly conclude that AI is equalizing the economy, while ownership gains are concentrating wealth and future income. The paper also shows that firms adopt AI selectively: automation is most attractive where labor costs are high and the technology can generate strong savings. Adoption therefore responds to the existing wage structure and can reinforce the value of capital.

The study does not forecast a complete Level 9 economy, and its results depend on assumptions about exposure, complementarity, and household portfolios. Its contribution is analytical clarity. Wage inequality, wealth inequality, and total welfare can move in different directions. Post-Labor Capitalism must therefore be evaluated through both flows of income and ownership of assets. Protecting wages alone may become insufficient if labor is no longer the main source of economic power.

Why it matters to Levels 1–9

It directly supports our claim that unemployment statistics alone are insufficient. Level 9 must track household assets, capital returns, and ownership distribution.

Limit or caution

The results are model-based and sensitive to assumptions about complementarity, adoption costs, and household portfolios.

Summary

This paper examines a far more advanced scenario than ordinary workplace automation: AI and robotics capable of performing essentially all economically useful work. The authors synthesize growth theory, automation models, and research on self-reproducing capital to ask what happens when human labor is no longer the main bottleneck in production.

The first critical threshold is full automation of production. If machines can build, maintain, and expand the capital stock with little human labor, capital becomes partly self-replicating. Increasing output no longer requires a proportional increase in workers. Under broad assumptions, this can raise the economic growth rate dramatically and cause labor’s share of income to collapse, breaking the relatively stable relationships that have characterized modern industrial economies.

A second threshold is automation of research and development. AI systems that can conduct science, engineering, and innovation may accelerate technical progress. However, the authors argue that automating R&D alone is not necessarily enough to create explosive growth if physical production, energy, natural resources, or other bottlenecks remain constrained. The interaction between automated ideas and automated production is what creates the most radical possibilities.

The effect on wages is ambiguous. Total output may expand enormously, which tends to raise incomes, while labor’s share may fall toward zero, which tends to lower wages. Which force dominates depends on returns to scale, the substitutability of machines and humans, natural-resource constraints, and the direction of technical change. Scarce nonreproducible inputs—land, minerals, energy locations, legal rights, or political control—may become much more valuable relative to labor and ordinary manufactured goods.

This is the closest core study to our technological definition of Level 9. It explains why a post-labor economy might differ fundamentally from earlier automation rather than merely continuing the same trend. It does not predict when transformative AI will arrive, and it leaves safety, politics, distribution, and social purpose largely outside the growth model. Its value is to identify the economic thresholds: production becomes post-labor only when machines can reproduce productive capacity and when the remaining human bottlenecks cease to bind. At that point, our current wage-centered institutions may no longer fit the structure of production.

Why it matters to Levels 1–9

This is the closest formal study to our technological definition of Level 9 and helps specify what must be true before “post-labor” is an accurate term.

Limit or caution

It deliberately analyzes extreme possibilities. Capability arrival dates, physical-resource bottlenecks, safety, politics, and distribution remain uncertain.

Summary

Korinek’s paper organizes AI policy according to capability levels rather than treating all AI as one technology. This is closely aligned with our Levels 1–9 method. The policies appropriate for systems that assist workers may be inadequate—or counterproductive—when AI can replace most labor, conduct research, control capital, and act autonomously across markets.

For the near term, the paper emphasizes familiar but urgent problems: measuring adoption and productivity, helping workers adjust, reforming education, maintaining competition, protecting privacy, and encouraging AI applications that complement people. Market incentives may favor labor-saving systems even when human-complementary innovations would create more widely shared benefits. Antitrust and intellectual-property policy also matter because compute, data, models, and platforms may concentrate economic power.

As capabilities advance, income distribution becomes more central. If labor income declines, social insurance and tax systems built around wages may weaken. Macroeconomic policy may face unusual combinations of rapid output growth, asset-price changes, falling labor shares, and concentrated demand. Education alone cannot solve a world in which machines can learn the newly taught skills. Governments may need new ways to distribute claims on production and stabilize consumer purchasing power.

The paper identifies eight broad policy challenges: inequality and income distribution; education and skills; social and political stability; macroeconomic policy; antitrust and market regulation; intellectual property; environmental implications; and global AI governance. International coordination becomes critical because advanced AI can alter military power, trade, tax competition, and the distribution of growth among countries. Energy and resource demands may also rise sharply.

The study’s principal contribution to our project is the staged policy logic. We should not select one permanent policy package today and assume it remains appropriate through Level 9. Worker training, for example, may be useful at Levels 2–4 but insufficient at Levels 7–9. Taxes on particular robots may be harmful early and irrelevant later. The paper does not settle the institutional details, but it gives us a map of the questions that must be revisited as capabilities cross thresholds. It supports a flexible transition framework rather than a single prediction or ideological prescription.

Why it matters to Levels 1–9

This aligns closely with our Levels 1–9 approach and can help convert the research library into a staged policy framework.

Limit or caution

The paper identifies policy domains more than it settles the political and administrative design of each institution.

Summary

This paper asks how governments can finance public services and redistribution when AI weakens the tax bases on which modern states depend. Advanced economies collect large amounts from labor income, payrolls, and human consumption. If AI replaces workers, wage and payroll revenue may fall precisely when displaced households need more support. The authors analyze public finance in two distinct stages of AI transformation.

In the first stage, human labor declines but people remain the principal consumers. The paper argues that consumption taxation may become more important because it can reach income generated by capital when that income is eventually spent. As labor distortions become less central, differential taxation of goods and services may also deserve reconsideration. Taxes should distinguish between final consumption and productive investment so that the system does not unnecessarily slow capital formation and innovation.

The second stage is more radical. Autonomous AI or AGI systems may produce most economic value and consume growing quantities of compute, energy, equipment, and other resources to expand themselves. If humans receive only a small fraction of output, taxing human consumption may no longer raise enough revenue or secure a meaningful claim on resources. The authors frame taxation of autonomous systems as an optimal “harvesting” problem: society must decide how much of the expanding AI-controlled resource base to claim for human use without destroying productive growth.

The paper evaluates concrete proposals—including robot taxes, compute taxes, token taxes, sovereign wealth funds, and windfall clauses—but does not endorse one simple instrument for every stage. A tax on robots may discourage useful early automation and be difficult to define. Compute or token taxes may be administratively easier in some settings but can distort production or move activity across borders. Public investment funds and windfall mechanisms can give society a claim on gains without taxing every machine transaction.

For the Capitalism without Labor, this study is foundational for Levels 6–9. It shows that taxation is not only about funding government; it determines how human beings retain access to output when autonomous capital dominates production. The analysis is theoretical and depends on future systems that do not yet exist at scale. Nevertheless, it directly addresses the institutional transition from a wage-tax state to a post-labor fiscal system.

Why it matters to Levels 1–9

This is foundational for Levels 6–9, where payroll taxes and employment-linked social insurance may no longer finance the state.

Limit or caution

The analysis is highly forward-looking and depends on assumptions about autonomous systems that do not yet exist at economy-wide scale.

Summary

This NBER overview introduces an emerging research field rather than presenting one isolated empirical result. Transformative AI is defined as AI capable of changing most economic production, potentially mastering the cognitive work humans perform and creating capabilities for tasks that do not yet exist. The overview argues that such systems could represent a change comparable to—or greater than—the Industrial Revolution.

The research agenda begins with growth. If AI automates production and innovation, output may expand much faster than historical experience suggests. Economists must then reconsider standard assumptions about stable labor shares, balanced growth, and the relative value of capital and labor. Reproducible inputs such as software, compute, and robots may become cheap, while scarce inputs such as land, energy, natural resources, legal authority, and trusted human control may become relatively more valuable.

A second cluster of questions concerns distribution and power. Transformative AI could reduce wages while raising returns to capital, concentrating wealth in firms or countries that control models, compute, data, and energy. It may also create new forms of autonomous economic agents, change the structure of firms, and alter competition. Traditional redistribution and tax systems may fail if labor and human consumption cease to be the dominant tax bases.

The field also studies geopolitical competition, information systems, governance, safety, and human well-being. Rapid technological progress can amplify national power differences and create pressure to deploy systems before institutions are ready. At the same time, economists can use scenario analysis to compare gradual automation with much faster transformative paths rather than betting on one arrival date.

For our Atlas, the overview confirms that the Level 9 question is becoming a legitimate research program, not merely science fiction. It also shows that no single paper answers the entire transition. Growth, firms, labor, competition, taxation, governance, and meaning require separate models that must later be connected. The source is accessible and useful as a gateway, but it reflects a young field with wide uncertainty and many preliminary papers. Its strongest message is methodological: transformative AI may be different enough that simple extrapolation from past automation is unsafe, so institutions should be stress-tested against several possible futures.

Why it matters to Levels 1–9

It is a useful gateway for readers before the more mathematical Level 9 studies and confirms that our central question is now becoming a recognized research field.

Limit or caution

It is an overview rather than a single empirical result, and the underlying field remains young and disputed.

Summary

Restrepo’s review brings together the main theoretical and empirical findings of modern automation economics. It begins with the task-based model, in which production is divided among activities that can be allocated to labor or capital. Automation expands the set of tasks performed by machines. This creates a displacement effect that lowers labor demand in the affected activities and tends to reduce labor’s share of value added.

The review emphasizes that productivity growth and displacement must be analyzed separately. Automation can lower costs and increase output, which may create demand for labor in complementary tasks and other industries. Yet an automation technology can be privately profitable while delivering only modest total productivity gains, especially if it replaces workers in tasks that people already perform efficiently. In that case, the displacement effect can dominate the broader benefits.

Restrepo surveys evidence from industrial robots, computerization, firm adoption, local labor markets, and cross-country industry data. The qualitative pattern is consistent across much of the literature: automation often reduces employment or wages for directly exposed workers, lowers labor’s share, and reallocates activity toward adopting firms and complementary occupations. Economy-wide employment effects are more mixed because price reductions, investment, and demand can create work elsewhere.

The review gives special importance to the creation of new tasks. Historically, technological progress generated new occupations and responsibilities that reinstated labor in production. Recent decades appear less favorable because task displacement accelerated while new-task creation and productivity growth were comparatively weak. The direction of innovation is therefore not neutral. Markets may invest heavily in labor-saving technologies because firms capture the wage savings, while underinvesting in technologies that make workers more capable.

For our project, this source is an efficient map of the entire task-based literature. It supports the view that automation should be evaluated by four separate measures: productivity, direct displacement, economy-wide demand, and new human tasks. It also reinforces a central uncertainty at higher levels: historical reinstatement may continue, or increasingly capable AI may automate new work faster than society can create it.

Why it matters to Levels 1–9

Useful orientation before reading the specialized studies.

Limit or caution

It is a synthesis, not an independent forecast.

Summary

This paper studies the early spread of AI through employer behavior rather than through predictions of technical capability. The researchers use a near-universe of online U.S. job vacancies beginning in 2010, with detailed information about occupations and required skills. They identify establishments that begin seeking AI-related talent and examine how the content of their other job postings changes afterward.

AI-adopting establishments sharply increase demand for workers with AI skills. More importantly, the task and skill requirements of non-AI vacancies also change. This suggests that AI adoption reorganizes work throughout the establishment rather than merely creating a small group of specialist positions. Some existing tasks become less prominent, while complementary technical, analytical, and managerial skills become more valuable.

Despite these visible changes inside adopting establishments, the authors do not find large, clearly measurable effects on aggregate employment or wages during the study period. AI adoption was still concentrated in a relatively small number of firms and occupations, so its economy-wide footprint remained limited. The paper therefore illustrates how technological transformation can begin beneath stable headline statistics.

The distinction between adoption and exposure is essential. Many jobs may contain tasks that AI could perform, but labor-market effects appear only when firms acquire systems, redesign processes, and change hiring. Job postings provide an early signal because employers alter the skills they seek before employment totals necessarily move. At the same time, postings reveal intended demand, not actual long-term staffing outcomes.

For the Levels 1–9 framework, this is a useful Level 1–2 study. It shows that the first signs of transition may be changes in job content, skill requirements, and reduced hiring for particular tasks rather than dramatic layoffs. It also suggests that traditional statistics can lag behind organizational change. By the time economy-wide employment effects become obvious, firms may already have spent years restructuring work around AI.

Why it matters to Levels 1–9

Relevant to the Level 1–2 “quiet transition.”

Limit or caution

Online vacancies do not capture every worker or firm.

Summary

This study asks how much of the widening U.S. wage structure since 1980 can be explained by changes in the tasks assigned to workers. The authors build a framework in which different demographic and education groups specialize in different tasks across industries. Automation and, to a lesser extent, offshoring displace workers from routine activities and reduce the demand for groups concentrated in those tasks.

The paper constructs measures of task displacement across industries and links them to changes in relative wages. Workers without college degrees—particularly men in production, operative, clerical, and administrative occupations—were more exposed to the tasks that disappeared. Their relative wages fell as those tasks were reassigned to machines or foreign labor. More highly educated workers were less exposed and were often complementary to the new technologies.

The authors estimate that task displacement accounts for a large share, roughly 50 to 70 percent, of the changes in the U.S. wage structure over several decades. Their analysis argues that automation is not merely correlated with inequality; it changes the demand for particular groups in a way that closely matches observed wage patterns. Other forces, including market power, declining unionization, and institutional change, may still matter, but they do not reproduce the same detailed pattern in the authors’ framework.

A striking implication is that large distributional changes do not require spectacular productivity growth. Technologies can be highly effective at replacing certain workers while adding relatively little to total output. Firms may adopt them because they save labor costs, even if the social productivity benefit is modest. This creates a combination of weak aggregate gains and severe losses for exposed groups.

For our project, the paper provides historical evidence that automation can reshape income distribution long before the economy approaches post-labor conditions. It also warns against treating “workers” as one category. The transition affects people differently according to their tasks, industry, education, age, gender, and location. A Level 1–5 model must therefore track distribution among groups, not only total employment and GDP.

Why it matters to Levels 1–9

Useful for measuring distributional damage before Level 9.

Limit or caution

The estimate depends on the authors’ task-displacement construction and period studied.

Summary

This paper addresses a puzzle that was especially visible before the generative-AI boom: AI systems were achieving impressive technical results, firms and investors were highly enthusiastic, yet measured productivity growth remained weak. The authors compare this situation with earlier general-purpose technologies and consider several explanations for the gap between expectations and statistics.

One possibility is that the optimism is simply wrong and AI will not have the broad economic impact its advocates expect. Another is mismeasurement: free digital services, rapid quality improvement, and intangible outputs may not be fully captured in GDP. The authors conclude that measurement problems exist but are unlikely to explain the entire productivity slowdown.

The explanation they find most persuasive is an implementation and restructuring lag. AI is not a plug-in replacement for one machine. Firms must build complementary assets, including data, software, training, new workflows, new products, and organizational systems. These investments take time, are risky, and are often recorded as current expenses. Productivity gains may therefore arrive only after businesses have learned how to redesign themselves around the technology.

Diffusion is also uneven. A small number of leading firms may achieve large gains while most businesses lack the skills, capital, or data to adopt effectively. Aggregate statistics remain weak until the technology spreads beyond the frontier. The same delay occurred with electricity and computers, whose largest benefits emerged after factories and offices were reorganized.

This study is the conceptual predecessor to the Productivity J-Curve. For our transition, it explains why Level 1 may last longer than technical demonstrations suggest. It also predicts a possible acceleration: once complementary investments are made and successful organizational designs become reproducible, productivity may rise more quickly. The paper does not determine whether those gains create or eliminate jobs, but it shows why short-run statistics are a poor guide to the eventual scale of transformation.

Why it matters to Levels 1–9

Helps prevent premature conclusions from early low productivity numbers.

Limit or caution

A paradox explanation is not proof that a future boom must occur.

Summary

This OECD report examines how AI changes the skills employers require, both in jobs that directly develop AI and in occupations that use it. It draws on occupational information, vacancy data, surveys, and the broader labor-market literature. The report avoids the assumption that every AI-exposed worker must become a programmer. Instead, it distinguishes specialized AI skills from the capabilities needed to work effectively alongside AI systems.

Demand for machine learning, data analysis, software, and digital literacy rises in firms adopting AI. However, technical skills are only one part of the change. Management, communication, teamwork, problem-solving, judgment, and domain knowledge can become more valuable when routine information processing is automated. Workers must understand when to trust an output, how to interpret it in context, and how to take responsibility for decisions.

The effects differ across occupations. In some jobs, AI reduces the importance of a skill because the system performs the task. In others, it raises demand for complementary expertise or allows a broader group of workers to perform previously specialized work. Skill requirements can therefore be upgraded, downgraded, or reorganized. Job titles may remain stable while the internal content changes substantially.

The report also identifies a training problem. Formal education systems change slowly, while firms often need occupation-specific adaptation. Employers may underinvest in training if workers can leave, and workers may be reluctant to invest when the useful life of a skill is uncertain. Access to training is frequently unequal, with highly educated employees receiving more support than workers most at risk of displacement.

For our Levels 1–5 analysis, the report helps specify the human side of transition. Early policy should not focus only on teaching AI development. It should strengthen complementary judgment, communication, responsibility, and the ability to supervise automated systems. At higher levels, however, the report’s training strategy may encounter limits if AI also becomes capable of the newly emphasized skills.

Why it matters to Levels 1–9

Useful for early transition and education policy.

Limit or caution

Skill demand today may not predict what remains scarce near Level 9.

Summary

This study measures how quickly generative AI entered everyday American life using nationally representative surveys. Unlike exposure studies, it asks people whether they actually use tools such as ChatGPT at work or at home, how frequently they use them, and for what activities. This provides an adoption baseline during the first years after public generative-AI systems became widely available.

By late 2024, nearly 40 percent of U.S. adults aged 18–64 reported using generative AI. About 23 percent of employed respondents had used it for work during the previous week, and roughly 9 percent used it every workday. This diffusion was faster than the early spread of personal computers and the internet, reflecting the fact that generative AI is delivered through existing devices and cloud infrastructure rather than requiring households to purchase entirely new hardware.

Usage was uneven. Younger, more educated, and higher-income workers adopted more rapidly, and use was especially common in computer, mathematical, and management occupations. Common workplace activities included writing, searching for information, obtaining instructions, summarizing, and generating ideas. Even workers in occupations with lower predicted exposure reported some use, showing that personal experimentation can precede formal company deployment.

The authors estimate meaningful time savings among users, but the aggregate productivity effect is smaller because many workers do not use the tools and many users apply them to only part of their work. Adoption does not reveal whether employers will later reduce staffing, expand output, or formalize the tools into automated workflows. Informal individual use is only an early stage.

For the Capitalism without Labor, the study shows that software AI can diffuse much faster than industrial technologies because the distribution infrastructure already exists. This may compress the early transition timeline. It also provides a reminder that widespread access is not the same as deep organizational integration. Level 2 begins when many workers use AI; later levels require firms to redesign production around it.

Why it matters to Levels 1–9

Relevant to estimating how quickly Level 1 can move toward Level 2.

Limit or caution

Self-reported use does not measure productivity or displacement directly.

Summary

This OECD chapter examines whether workers and their representatives can influence the way AI is introduced. Its starting point is that the same technical system can produce very different workplace outcomes. AI may remove dangerous tasks, improve decisions, and support training, or it may intensify surveillance, work pace, and insecurity. Implementation is therefore a governance question as well as an engineering decision.

Social dialogue can begin before procurement, when firms decide which problems AI should solve. Worker representatives may identify practical risks that designers and managers miss, including biased evaluation, unrealistic performance targets, privacy concerns, and the loss of tacit knowledge. Consultation can also clarify which tasks will change and what training is required.

Collective bargaining can establish rules for data use, monitoring, transparency, retraining, redeployment, and the sharing of productivity gains. It can help employees trust systems that might otherwise appear to be tools of unilateral control. In some settings, unions and works councils have negotiated protections or participated in technology committees. However, many worker organizations lack AI expertise, and collective bargaining coverage is weak in parts of the economy.

The OECD does not claim that dialogue eliminates displacement or prevents firms from automating. Its position is that participation can improve the quality and fairness of the transition and help businesses implement systems more effectively. It may also shift AI development toward complementing workers rather than merely reducing headcount.

For our project, this source is most relevant to Levels 1–4, while employees still possess significant bargaining power and companies still depend on their cooperation and knowledge. It highlights a time-sensitive issue: worker influence may be strongest before automation is complete. By Levels 7–9, traditional unions built around employment may have far less leverage, requiring new institutions of citizen, consumer, or ownership representation.

Why it matters to Levels 1–9

Relevant to Levels 1–5 and company-level transition rules.

Limit or caution

Collective bargaining coverage and effectiveness vary greatly among countries.

Summary

This paper investigates one of the main historical answers to automation: the economy creates work that did not previously exist. The authors use U.S. Census occupational data, job titles, patents, and textual information to identify new forms of work between 1940 and 2018. Instead of counting only entirely new occupations, they examine new specialties and titles appearing inside broader occupational categories.

New work is a substantial part of modern employment. Technological advances create roles associated with designing, operating, marketing, maintaining, and applying new products and systems. Rising incomes also generate new services and specialties as consumers demand health, education, entertainment, personal care, and other activities. New work therefore emerges from both the supply of new technology and changing demand.

The content is not evenly distributed. New technology-related work often appears first in higher-skill occupations, while demand-driven new services can appear across the wage distribution. Over time, some specialties become standardized, diffuse to a broader workforce, or are themselves automated. The creation of new work can support labor demand, but it does not guarantee that displaced workers can move directly into the new positions.

The historical record supports the reinstatement mechanism in task-based theory. Much of today’s employment consists of job titles that were absent in earlier censuses. At the same time, the paper does not prove that future new work will keep pace with AI. Past technologies left large areas of cognitive and physical activity for people; Level 9 is defined by the possibility that machines can enter almost all of them.

For our project, this study tells us what to measure between Levels 1 and 8: not only occupations created and destroyed, but the origin, skill level, pay, accessibility, and durability of new tasks. A transition can look successful in total job numbers while creating work that is inaccessible to displaced employees or vulnerable to rapid subsequent automation.

Why it matters to Levels 1–9

Directly relevant to the user’s expectation that people will continue to seek occupations.

Limit or caution

Past creation of work may not continue if AI can absorb new tasks faster than humans create them.

Summary

Acemoglu and Johnson present a broad historical argument that technological progress does not naturally create shared prosperity. New machines can increase productivity while concentrating wealth and political power. Whether gains spread depends on the direction of innovation, the institutions governing adoption, and the ability of workers and citizens to influence how technology is used.

The authors contrast technologies that primarily automate and monitor people with technologies that make workers more productive and create new capabilities. Elites often promote a “productivity bandwagon” story: once output rises, everyone will eventually benefit. Historical episodes show that this outcome is not automatic. Early industrialization produced harsh conditions and weak wages before political reform, unionization, education, and new technologies helped workers share in growth.

The book ranges across agriculture, industrial machinery, communications, mass production, and digital platforms. It argues that centralized control over information and production can reinforce existing hierarchies. Modern digital systems often automate routine work, monitor employees, and create winner-take-all markets rather than broadly expanding human expertise. This reflects incentives and ideology, not an unavoidable property of computing.

The authors advocate a more “machine-useful” direction for innovation: technologies should extend human judgment, create new tasks, improve public services, and distribute decision-making more broadly. They emphasize competition policy, worker voice, public research, taxation, and institutional reforms that change which innovations are profitable and socially rewarded.

For the Capitalism without Labor, Power and Progress supplies the political-economy layer missing from many formal models. Ownership and income distribution are not only technical tax questions; they reflect struggles over who controls technology and sets its purpose. The book is skeptical that concentrated private incentives alone will guide society toward a desirable Level 9, but it does not reject markets or entrepreneurship. Instead, it argues that democratic institutions must shape the technological path if progress is to benefit the majority.

Why it matters to Levels 1–9

Useful for framing the transition as a political choice rather than a technical inevitability.

Limit or caution

It is a historical argument and policy thesis, not a Level 9 model.

Summary

Daniel Susskind examines the possibility that technological unemployment becomes structural rather than temporary. He does not argue that every job disappears at once. Instead, he describes “task encroachment”: machines gradually take over a growing range of activities, reducing the amount and variety of paid work available to people even as new tasks continue to appear.

The book challenges the traditional response that workers can always retrain. Education may help individuals compete during early stages, but it cannot solve a situation in which machines become capable of performing the newly taught skills as well. The long-run problem shifts from preparing everyone for new jobs to deciding how income, power, and meaning are organized when paid labor is no longer universally necessary.

Susskind gives strong attention to distribution. The owners and controllers of productive technology may receive an increasing share of income, while workers lose their main claim on economic output. He argues for a larger role for the state in distributing prosperity and maintaining social participation. His proposals include forms of income support, taxation, and what he describes as conditional arrangements that connect public support with socially valued activity.

The book also treats work as more than a paycheck. Jobs provide status, structure, community, and a sense of contribution. A society that replaces wages without creating alternative institutions of purpose could remain psychologically and politically unstable. Education, culture, civic institutions, volunteering, care, and community participation may become central parts of a post-work society.

For our project, this source directly supports the distinction among job, work, and occupation. It helps us ask what people will do at Level 9, not merely what they will consume. Its approach is broader and less mathematically formal than the economic papers, but it connects technological unemployment to political power and human meaning. It is therefore valuable when designing the social side of Post-Labor Capitalism.

Why it matters to Levels 1–9

Relevant to our job–work–occupation distinction.

Limit or caution

The book predates the current generative-AI wave and offers a normative argument rather than tested institutional design.

Summary

Brynjolfsson and McAfee argue that digital technologies represent a new economic era because they improve exponentially, can be reproduced at near-zero marginal cost, and can be combined into new products and processes. These characteristics allow innovation to spread rapidly and create forms of abundance that were difficult under purely physical production.

The authors describe the benefits as the “bounty”: greater productivity, lower prices, improved quality, and access to services that once required expensive expertise. Digital tools can augment workers, create new industries, and allow small teams to reach global markets. This is the optimistic side of the second machine age.

The “spread” is the unequal distribution of those gains. Digital markets often reward a small number of top performers, platform owners, innovators, and people with complementary skills. Software can serve millions of customers without proportional employment, creating superstar firms and winner-take-most outcomes. Routine and middle-skill workers face pressure, while capital owners capture a growing share of value.

The book recommends investment in education, entrepreneurship, infrastructure, research, organizational innovation, and policies that make it easier for people to create and adapt. It was written before modern generative AI, but it anticipated the combination of rapid technological improvement, high productivity potential, and widening inequality that now shapes the debate.

For the Capitalism without Labor, the book is an important early bridge from the computer age to the AI era. It explains why digital production can grow without matching employment growth and why entrepreneurship may become accessible to smaller teams. At the same time, its proposed adaptation policies largely assume that humans remain economically complementary to machines. Higher levels may require extending the analysis from skill adaptation to ownership and income systems.

Why it matters to Levels 1–9

Useful historical context for early Atlas chapters.

Limit or caution

Written before modern foundation models and general-purpose robotics.

Summary

Keynes wrote this essay during the Great Depression but deliberately looked a century ahead. He argued that technological progress and capital accumulation could eventually solve what he called the “economic problem”: the struggle to secure enough food, shelter, and material necessities. He expected living standards to rise several times over and imagined that people might need to work only about fifteen hours a week.

The essay introduces the phrase “technological unemployment” to describe labor displaced because society discovers ways to economize on workers faster than it discovers new uses for them. Keynes viewed this as a temporary adjustment problem on the path to abundance rather than a permanent catastrophe. In his expectation, economic growth would eventually create enough wealth to reduce the necessity of labor.

He distinguishes absolute needs, which can be satisfied, from relative needs, which arise from the desire to feel superior to others and may expand indefinitely. Even in an affluent society, competition for status could preserve long hours and consumption. The technological solution to scarcity would therefore not automatically solve the cultural problem of knowing when enough is enough.

Keynes believed the greatest challenge would be learning how to use freedom and leisure. People formed by centuries of economic necessity might struggle to find purpose when work no longer organizes life. He expected society to value art, relationships, contemplation, and activities pursued for their own sake, but recognized that the transition could be psychologically difficult.

For our project, this essay is an early statement of the Level 9 social question. Its quantitative forecast was imperfect and its assumptions reflect the society of 1930, but its conceptual distinction remains powerful: technological capacity can reduce necessary labor without teaching people how to live meaningfully. Post-Labor Capitalism must therefore address both purchasing power and occupation beyond compulsory employment.

Why it matters to Levels 1–9

Relevant to the human-purpose side of Level 9.

Limit or caution

It underestimated some new wants and could not anticipate modern AI, services, or environmental constraints.