Reprint Statement

This article is reprinted from Caijing Magazine, original author: Zhuo Xian, editor: Su Qi. Copyright belongs to the original author and publisher. Reprinted for educational and exchange purposes only.

China's "AI+" action plan emphasizes large-scale application of technology, which not only shortens the return on investment cycle for technological innovation but also helps increase employment by creating AI application scenarios in production, consumption, and distribution.

The Relationship Between Economic Growth and Employment is Changing

Before the Industrial Revolution, during the long agrarian era, low technological progress brought low-speed growth, corresponding to low population growth and low employment growth. Economic growth was almost equivalent to the growth of agricultural employment.

The Industrial Revolution broke through the constraints of energy and power and the existing combinations of production factors, greatly expanding humanity's production frontier. Industrialization and urbanization promoted each other based on economies of scale. Industrial product prices fell as productivity improved, and wage levels rose as productivity improved. Large-scale production and large-scale consumption formed a positive cycle, and industrial blue-collar jobs increased rapidly.

The modern corporate system expanded the scope of social division of labor and collaboration. Numerous production links originally completed within a single enterprise (such as logistics, marketing, and legal consulting) became independent specialized enterprises, forming a vast network of intermediate inputs and services. While improving economic efficiency, this also led to a large number of knowledge-based white-collar positions in the corporate hierarchy. After the popularization of personal computers in the 1980s, information processing positions such as accounting, secretarial work, and analysts grew rapidly.

The marketization of household labor is another important engine for job creation. As women entered the labor market on a large scale, labor originally performed unpaid within households was transformed into marketized services in national economic accounting. Jobs in domestic services, catering, education, entertainment, and other life service industries were continuously created.

For most of the 20th century, "economic prosperity means full employment" was a social perception shaped by industrial civilization, and became the narrative logic and psychological foundation of many current business models and social systems.

The several "jobless growth" experiences of developed economies around the turn of the 21st century began to challenge this consensus. Initial research suggested that "jobless growth" occurred after economic crises, mainly from new enterprises increasing equipment investment, and was a cyclical abnormal phenomenon that did not form a structural change in the employment-growth relationship. However, subsequent research showed that the disappearance of routine cognitive and physical work did not happen gradually, but was concentrated during economic recessions. Enterprises used crises as a concentrated "cleansing mechanism," permanently eliminating middle-skill positions that could be automated. When the economy recovered, these jobs would not return. Although the service industry eventually absorbed most of the labor force, it was at the cost of sacrificing wage growth and job stability.

Comprehensive reviews of domestic and international literature on AI's impact on employment in recent years show that artificial intelligence has not caused large-scale unemployment. Many studies also found that although unemployment rates in high-AI-exposure industries are indeed rising, unemployment rates are rising even faster in lower-exposure occupations. One possible explanation is that workers in high-AI-exposure industries have higher education levels, stronger re-employment capabilities, and are therefore less affected. The few studies proving higher unemployment rates among high-AI-exposure populations mainly use large language models to assess the risk of various occupations being replaced by artificial intelligence—essentially "artificial intelligence telling us that artificial intelligence is exacerbating unemployment"—but the statistical significance is not high.

Although the impact on total employment is not obvious, in the current AI era, the relationship between employment and growth has shown some new trends, which can be summarized as three aspects of "decoupling."

First, Decoupling of Employment and Investment

In the industrial era and service economy era, whether infrastructure investment or machinery and equipment investment, both brought considerable direct and indirect employment. In the AI era, technology companies are deepening capital at an unprecedented speed, but the job creation effect is declining.

Unlike the last round of internet investment boom, the expansion model of the AI era has shifted from "light assets, heavy manpower" to "heavy capital, heavy computing power," relying on high-density investment in physical infrastructure such as data centers and energy networks. Microsoft, Amazon, Google, and Meta's total capital expenditure in 2025 is expected to reach $400 billion, a scale exceeding the annual GDP of many medium-sized countries. However, technology companies are simultaneously implementing human capital tightening strategies, cutting hundreds of thousands of jobs and freezing entry-level recruitment for graduates. Unusually, these behaviors occur against a backdrop of technology companies' stock prices hitting record highs and strong revenue growth, reflecting corporate decision-making thinking that cuts labor costs to release funds for computing infrastructure investment.

Second, Decoupling of Technological Progress and Human Capital Improvement

In the past, improvements in labor productivity came both from technological contributions embedded in capital and machinery, and from human capital improvements in the "learning by doing" process. In the AI era, improvements in labor productivity are more likely to come from the denominator of this indicator (i.e., "labor force scale") declining, with the speed of human capital improvement far behind the speed of AI technological progress.

On the one hand, the accumulation path of human capital "learning by doing" is narrowing. In the past, university graduates accumulated experience through basic work and gradually grew into senior talents. Now, AI is increasingly competent at junior analyst, junior programmer, and junior copywriter jobs, and recruitment demand for some positions targeting fresh graduates is declining. For example, traditional law firm models rely on large numbers of junior lawyers for document review and legal retrieval work. Now AI can complete this work in seconds, but divorce and other case demands will not rise due to AI technology development, leading law firms to significantly reduce junior lawyer recruitment. This not only leads to rising youth unemployment rates but may also cut off the ladder for human capital improvement in many types of positions for a long time—if enterprises no longer recruit junior employees, where will future senior experts come from?

On the other hand, in the competition between technology and education, the linear accumulation speed of human capital cannot keep up with the exponential speed of technological evolution. A major prescription for employment challenges in the AI era is lifelong education. However, changes in education models are not a panacea in the face of AI technology progress. Most workers' human capital accumulation speed can no longer catch up with the evolution speed of machine intelligence. For example, when universities just start offering "prompt engineering" courses, the latest models may no longer need prompt optimization.

Third, Decoupling of Worker Wages and Productivity Growth

Research on the U.S. employment market shows that the decoupling of labor productivity and real wages has been ongoing since the 1970s, and the accelerated application of AI may widen this gap. In the AI era, AI drives the routinization of non-routine cognitive tasks such as junior code writing, legal document drafting, and basic financial analysis. Excess profits in high-efficiency sectors are more converted into capital returns and salary growth for a small number of core talents. Auxiliary position workers remaining in high-efficiency sectors not only tend to decrease in number, but because their human capital contribution is less than AI, wage growth will not keep pace with industry productivity improvements.

The traditional "Baumol-style" productivity sharing mechanism fails. Baumol's "cost disease" theory points out that excess value created by high-productivity sectors such as manufacturing spills over to low-productivity-growth sectors such as healthcare, nursing, and entertainment through labor market competition (competing for scarce labor) or institutional arrangements (union negotiations, minimum wage, etc.), thereby achieving universal wage increases across society. This cross-sector wage transmission mechanism maintained relative equilibrium in the labor market and became the main channel for practitioners in low-efficiency sectors to share prosperity dividends. In the AI era, because high-efficiency sectors no longer need more positions, they do not need to continuously raise wages to maintain their workforce, and thus cannot "demonstrate" higher wages across society through the "wage demonstration effect." When middle-skill workers replaced by AI (such as clerks, translators, junior coders) flow to low-productivity-growth service industries (such as ride-hailing, delivery, basic nursing), labor supply exceeds demand, and the mechanism of low-efficiency sector worker wages rising with high-efficiency sector wages is blocked.

Declining AI costs suppress the "hard ceiling" of human wage increases. For a large number of tasks based on rules, logical analysis, information synthesis, and pattern recognition, AI provides nearly unlimited supply. The scarcity of human capital in these fields is broken, and the market price of related skills tends to decline. AI technology is essentially energy-intensive. If the marginal cost of intelligence eventually converges to energy costs, and energy costs continue to decline with technological innovations such as controllable nuclear fusion, high-altitude wind power generation, and space photovoltaics, the wage ceiling for humans completing existing tasks faces continuous downward pressure. For example, in a certain task, when AI deployment costs drop to $5 per hour, workers originally only doing this single task can never earn more than $5, no matter how much their productivity improves.

---

Social Insurance Systems Based on Stable Employment Face Challenges

Based on different assumptions about the timing, speed, and scope of AI's substitution and creation effects on employment, different institutions' "crystal balls" vary greatly in predicting AI's impact on future employment. For example, since 2020, the World Economic Forum has made three consecutive opposite judgments on whether artificial intelligence increases employment, with a gap of 92 million in predictions of global net job additions and losses over the next five years. Compared to changes in total employment, this article focuses more on the challenges that structural changes in employment in the AI era pose to social security.

Modern social insurance systems are products of the large-scale industrialization era. Whether public pension insurance and medical insurance, or unemployment insurance, work injury insurance, or maternity insurance, their original intention was to achieve socialized dispersion of "worker employment interruption risk." Therefore, social security system design is strongly linked to employment contributions, and its continued operation relies on three pillars: employment growth brought by demographic dividends, standardization of labor relations formed by large-scale industrial production, and wage income growth driven by productivity improvements. It was the historical convergence of these three conditions in the 20th century that made social insurance systems financially feasible and politically operable, becoming an important institution for states to manage social risks.

First Pillar: Favorable Population Structure

Favorable population structures provide the actuarial foundation for social insurance. Under social insurance systems, population growth itself is transformed into a special asset class, with intergenerational transfer payments producing an implicit "biological rate of return" that can even exceed monetary capital accumulation. If an economy's population growth rate (n) plus real wage growth rate (g) is greater than the market real interest rate (r), introducing a pay-as-you-go social insurance system will increase total social welfare.

In the decades of the "golden age" after World War II, the baby boom made this "return without capital" a reality. Participating in social security was not just a mandatory burden, but a better investment behavior than private savings. Favorable population structures established a socially consensual social security intergenerational contract, achieving the transfer of retirement risk management from dispersed families to centralized social provision.

Second Pillar: Long-term Stable Employment Relationships

Unlike social relief based on means testing, modern social security systems emphasize the equivalence of rights and obligations—that is, benefit levels are strictly linked to contribution history. This design aims to maintain workers' decent living standards after retirement. Long-term stable employment relationships give workers clear, coherent income streams, ensuring the possibility of linking "retirement benefits" to "labor contributions."

Highly organized employment relationships not only created a stable middle class but also made workers' income transparent, calculable, and easy to deduct. This transformed the modern enterprise system into an extension of state capacity, making enterprises agents for the state's wage tax (fee) extraction, improving the administrative efficiency of social security fund collection and expanding its coverage.

Third Pillar: Synchronized Growth of Worker Wages and Productivity

Synchronized growth of wages and productivity ensures endogenous expansion of social security contribution bases. With population structures and collection mechanisms determined, the improvement of social security benefit levels and fund solvency fundamentally depends on the growth rate of the contribution base. Even with population aging, when n declines or even becomes negative, if real wage growth rate g maintains high growth, social security welfare levels can naturally improve with increases in total social wealth.

In the 30 years after World War II, Western countries experienced a golden period of productivity growth. High unionization rates ensured that productivity improvements were converted into wage growth, forming a virtuous cycle of broadly shared productivity gains. The compound growth formed by population dividends叠加 productivity dividends allowed each generation to support the previous generation to live better than when they were young by contributing only a small portion of their income.

---

Modern social insurance systems are institutional arrangements designed by human society through rational design to manage industrialization risks. They successfully internalized three specific macro-historical conditions as parameters for system operation, increasing social cohesion and improving economic and social stability. However, since the late 20th century, population aging has shaken the actuarial logic of the first pillar, and the second and third pillars also face challenges in the leap-forward development of artificial intelligence technology.

The impact of population aging on the first pillar has been discussed extensively and will not be repeated here. However, it should be pointed out that the impact of aging on social insurance systems is gradual and predictable, while the progress of artificial intelligence is non-linear and exponential, and may pose challenges to the second and third pillars of existing social security models that are faster, broader, and larger in scale.

First, Artificial Intelligence Will Change Industrial Civilization's Production Organization Models and Enterprise Forms

This will fragment original formal employment relationships and shake the second pillar.

On the one hand, artificial intelligence reduces market transaction costs and promotes the gig economy for knowledge workers.

If markets are effective resource allocation mechanisms, why do enterprises exist? Coase's answer is: market transactions have search, bargaining, contracting, and supervision costs. When an enterprise's internal organizational costs are lower than external market transaction costs, enterprises emerge and expand. With the application of AI technology in labor market platforms, the transaction cost of "hiring by task" becomes negligible compared to "hiring by job." The basic unit of work gradually shifts from a "job" (Job)—a bundle of long-term, vague task collections—to a "task" (Task)—single, clear, short-term deliverables—or even the so-called "Coase singularity" appears.

Under the Coase singularity, large amounts of core tasks originally belonging to enterprises can be outsourced, or even "one-person companies" appear. Workers originally employed long-term and stably by enterprises become outsourced personnel. Financial reports from global freelancer platforms such as Upwork and Fiverr show that large enterprises are systematically replacing full-time employees with high-skill freelancers. If "enterprises"—the core nodes of social security collection—are replaced by "transaction networks" of knowledge-based tasks, the possibility of more office white-collar positions shifting from fixed employment to gig work will increase.

On the other hand, artificial intelligence reduces enterprise internal coordination costs, potentially forming a "middle management collapse."

In traditional enterprises, middle managers' core functions are information transmission, task allocation, and process monitoring. AI agents begin to execute complex workflows without continuous human intervention and complete these coordination tasks at extremely low costs. This may lead to flattening of enterprise organizational structures, where senior leaders can directly supervise more business units, and middle managers responsible for coordinating tasks and information processing are no longer indispensable. Gartner predicts that by 2026, 20% of organizations will use AI to flatten organizational structures, and more than half of middle management positions will no longer be needed.

Both aspects above will cause the gig economy to develop from current construction, manufacturing, and life service industries such as food delivery and express delivery to knowledge white-collar-dominated producer service industries, creating larger-scale non-long-term employment relationships. This leads to declining employer responsibility for social insurance contributions and rising individual contribution responsibility and risk exposure for workers.

Furthermore, if Large-scale AI Capital Deepening Continues in its Current Form

The tilt of national income distribution toward capital owners and a small number of high-skill workers will shake the third pillar.

Artificial intelligence may make it difficult for middle-income groups' wage income to synchronize with productivity improvements. The main source of social insurance system funds is the vast middle-income group. Unlike previous industrial revolutions that mainly replaced blue-collar physical labor, generative AI accelerates the routinization of non-routine cognition, turning mid-to-high-level cognitive abilities into industrially replicable services. It mainly impacts white-collar workers with higher education engaged in cognitive work—a group with stable jobs, higher wages, and high compliance contribution rates.

Declining labor compensation proportions will lead to relative declines in social security tax bases. Data from the OECD and International Labour Organization both show that in industries with the highest digitalization, the share of labor income in value added is accelerating downward. This means that dividends from technological progress flow more to capital owners who own algorithms, data, and computing power. Since high-income earners have caps on contributions to public basic pension insurance, medical insurance, and unemployment insurance, further income growth for this group contributes almost nothing to social security funds. If AI-era capital deepening leads to reduced labor income shares, especially middle-income group income shares, the proportion of social security tax bases to total economic output will decline, and economic growth will not be converted into synchronized growth of social security funds.

---

Building Employment-friendly Development Models in the AI Era

Technology itself is neutral, but technological innovation does not naturally lead to human welfare. If the purpose of artificial intelligence is to improve human potential and quality of life, rather than "how to replace people with machines," the challenges described above will be readily solved, and technological dividends can compensate for disappearing population dividends. For example, the European Medical Technology Industry Association estimates that widespread AI application in healthcare could save European healthcare systems €170-210 billion annually, with wearable AI devices alone saving approximately €50 billion annually, directly reducing social security fund pressure on drug procurement.

Another important way to solve the pension crisis is to increase contribution years. AI technology can eliminate physiological and cognitive barriers for elderly people participating in the labor market, allowing older employees to focus on high-value work requiring judgment, empathy, and complex decision-making, reducing work fatigue. This enables elderly people to choose a "gradual retirement" approach transitioning from full-time to part-time work, rather than suddenly cutting off income sources.

However, at least four factors currently make the direction of artificial intelligence innovation unfavorable to employment and social security:

First, the capital-driven "Turing Trap." Stanford University's Erik Brynjolfsson proposed the concept of the "Turing Trap," pointing out that current AI R&D over-focuses on "thinking and acting like humans," developing "human-like intelligence" rather than enhancing human capabilities. This is the result of capital-driven innovation responding to scarcity. Price, as a signal of scarcity, directs the direction of technological change, making innovation tend to replace factors that are large in scale and high in price. In developed economies, this directs innovation toward replacing high-cost labor.

Second, geo-economics promotes labor-saving innovation routes. In recent years, under the influence of geo-economics, developed economies have promoted industrial reshoring but face severe shortages of skilled labor. To avoid uncertainties in cross-border investment, immigration policy, and tariff policy, enterprises have shifted the focus of technology investment toward "labor-saving" directions.

Third, endless demand in the bit world exacerbates scarcity in the atom world. AI technology innovation cannot directly break atomic scarcity. Physical constraints on land, freshwater, lithium, cobalt, and other key minerals still exist. Economic growth scarcity has shifted to energy, environmental capacity, and key raw materials. From an employment perspective, these are all labor-thin fields. If accelerated development occurs, it may also cause artificial intelligence to compete with human welfare for scarce resources.

Fourth, limitations of AI4Science innovation. A study analyzing 67 million papers in six fields—biology, chemistry, geology, materials science, medicine, and physics—points out that although AI tools improve individual scientists' output, they lead to convergence in scientific research topic selection. That is, scientists tend to study data-rich fields that AI can easily process, while data-scarce or marginal fields that AI cannot model are neglected. This tendency may narrow the breadth of scientific discovery and reduce the possibility of breakthrough innovations opening up human needs and employment space.

Technological progress has path dependence. Once a certain technological paradigm becomes dominant, the entire society's engineering capabilities, infrastructure, and cognitive habits will be built around it and self-reinforce, "locking in" development patterns on specific tracks. The "14th Five-Year Plan" proposal suggests "building employment-friendly development models" and explicitly calls for "improving employment impact assessment and monitoring and early warning" to address "the impact of new technology development on employment." This is the unity of high-quality development and high-quality full employment, and is of great significance for guiding the development of artificial intelligence technology in the right direction.

Unlike the U.S., which bets most incremental innovation resources on AI's training and inference layers, China's proposed "AI+" action plan emphasizes large-scale application of technology, with innovation resources more evenly distributed across AI's training layer, inference layer, and application layer. This not only shortens the return on investment cycle for technological innovation but also helps increase employment by creating AI application scenarios in production, consumption, and distribution. Moreover, China's labor costs are far lower than the U.S., and the benefits of AI replacing labor are not high, leaving more room to promote AI technology "for good" development through public policy.

In addition to already deployed conventional policies, this article proposes several policy recommendation directions for discussion:

On "Robot Taxes"

Because some countries provide tax deductions or accelerated depreciation policies for automation equipment, while levying high wage taxes including social security on labor, this actually subsidizes the behavior of replacing humans with AI technology. Although many studies propose robot taxes, there is currently no policy practice in various countries. The South Korean government, often mistakenly called implementing the "world's first robot tax," does not directly tax robots but reduces tax credits for enterprise investment in automation equipment.

Robot taxes can theoretically internalize social costs of AI development (such as unemployment) and slow down excessive job substitution, but face definitional difficulties in operation. For example, what is a robot? Does Excel improved by AI technology need to be taxed? A more likely operational path is to implement differentiated tax rates based on AI technology types: providing tax credits for "labor-augmenting" technologies such as exoskeletons and augmented reality glasses that assist workers; not providing tax benefits or moderately taxing technologies that purely replace labor.

On "Tax-Fee Coordination" Social Security Financing Methods

Unlike Germany, France, and other continental European countries that mainly rely on employer and employee contributions, countries like Denmark choose paths with general taxation as the main funding source, where contributions account for a smaller proportion of social security financing. Japan is one of the world's most aging countries. In 2019, it increased the consumption tax rate from 8% to 10%, explicitly dedicating increased consumption tax revenue to pension, medical, and nursing social security expenditures.

Although Denmark's social security financing structure and Japan's social security reform were not originally aimed at AI shocks, "tax-fee coordination" social security financing methods can allow wealth dividends created by AI to flow back to social security networks, alleviating impacts on the three pillars of social security. As for specific tax types, from policy practices in some countries, value-added tax (or consumption tax), environmental tax, and capital gains tax are options. Some research institutions have also proposed levying AI "excess profit taxes."

On Sovereign AI Infrastructure

If AI computing power becomes future currency as some studies suggest, then mastering AI infrastructure means mastering future seigniorage. Building "sovereign AI infrastructure" is not only a national security issue but may also become a new channel for social security financing. The UK, France, Canada, and Singapore are investing in building state-owned "national research clouds" or sovereign AI computing clusters. By holding core computing infrastructure through national investment, governments can directly capture economic rents generated by future AI.

After AI large-scale commercial application, this "AI dividend" can play a role similar to Norway's current oil fund, directly injecting into social security systems, achieving a shift from "taxing labor" to "AI dividends" alongside taxing labor, allowing social security systems to share in capital appreciation brought by AI.

On Human Capital Accumulation Methods in the AI Era

A study by European think tank Bruegel found that in AI-related job postings, the proportion mentioning university degrees declined by 23%, while the proportion mentioning specific skills rose sharply. In basic education and higher education stages, due to the shortened half-life of specific professional backgrounds and skills, education must shift toward cultivating "meta-cognitive" abilities, critical thinking, and interdisciplinary system integration capabilities.

In terms of youth employment, as AI takes over junior work, the original "learning by doing" human capital path narrows. New graduate internship incentive mechanisms must be designed. Consider using fiscal funds to subsidize wages or pay social security on behalf of young people entering the workplace, encouraging enterprises to hire young people and carry out human-machine collaboration growing together with AI in work.